Sei sulla pagina 1di 593

Contents i

A Course in Mathematical Statistics


Second Edition
ii Contents

This Page Intentionally Left Blank


Contents iii

A Course in Mathematical Statistics


Second Edition

George G. Roussas
Intercollege Division of Statistics
University of California
Davis, California

ACADEMIC PRESS
San Diego London Boston
New York Sydney Tokyo Toronto
iv Contents

This book is printed on acid-free paper.

Copyright 1997 by Academic Press

All rights reserved.


No part of this publication may be reproduced or transmitted in any form or by
any means, electronic or mechanical, including photocopy, recording, or any
information storage and retrieval system, without permission in writing from the
publisher.

ACADEMIC PRESS
525 B Street, Suite 1900, San Diego, CA 92101-4495, USA
1300 Boylston Street, Chestnut Hill, MA 02167, USA
http://www.apnet.com

ACADEMIC PRESS LIMITED


2428 Oval Road, London NW1 7DX, UK
http://www.hbuk.co.uk/ap/

Library of Congress Cataloging-in-Publication Data

Roussas, George G.
A course in mathematical statistics / George G. Roussas.2nd ed.
p. cm.
Rev. ed. of: A first course in mathematical statistics. 1973.
Includes index.
ISBN 0-12-599315-3
1. Mathematical statistics. I. Roussas, George G. First course in mathematical
statistics. II. Title.
QA276.R687 1997 96-42115
519.5dc20 CIP

Printed in the United States of America


96 97 98 99 00 EB 9 8 7 6 5 4 3 2 1
Contents v

To my wife and sons


vi Contents

This Page Intentionally Left Blank


Contents vii

Contents

Preface to the Second Edition xv

Preface to the First Edition xviii

Chapter 1 Basic Concepts of Set Theory 1

1.1 Some Definitions and Notation 1


Exercises 5
1.2* Fields and -Fields 8

Chapter 2 Some Probabilistic Concepts and Results 14


2.1 Probability Functions and Some Basic Properties and Results 14
Exercises 20
2.2 Conditional Probability 21
Exercises 25
2.3 Independence 27
Exercises 33
2.4 Combinatorial Results 34
Exercises 40
2.5* Product Probability Spaces 45
Exercises 47

vii
viii Contents

2.6* The Probability of Matchings 47


Exercises 52

Chapter 3 On Random Variables and Their Distributions 53

3.1 Some General Concepts 53


3.2 Discrete Random Variables (and Random Vectors) 55
Exercises 61
3.3 Continuous Random Variables (and Random Vectors) 65
Exercises 76
3.4 The Poisson Distribution as an Approximation to the Binomial
Distribution and the Binomial Distribution as an Approximation to the
Hypergeometric Distribution 79
Exercises 82
3.5* Random Variables as Measurable Functions and Related Results 82
Exercises 84

Chapter 4 Distribution Functions, Probability Densities, and Their


Relationship 85

4.1 The Cumulative Distribution Function (c.d.f. or d.f.) of a Random


VectorBasic Properties of the d.f. of a Random Variable 85
Exercises 89
4.2 The d.f. of a Random Vector and Its PropertiesMarginal and
Conditional d.f.s and p.d.f.s 91
Exercises 97
4.3 Quantiles and Modes of a Distribution 99
Exercises 102
4.4* Justification of Statements 1 and 2 102
Exercises 105

Chapter 5 Moments of Random VariablesSome Moment and


Probability Inequalities 106

5.1 Moments of Random Variables 106


Exercises 111
5.2 Expectations and Variances of Some R.V.s 114
Exercises 119
5.3 Conditional Moments of Random Variables 122
Exercises 124
Contents ix

5.4 Some Important Applications: Probability and Moment


Inequalities 125
Exercises 128
5.5 Covariance, Correlation Coefficient and Its Interpretation 129
Exercises 133
5.6* Justification of Relation (2) in Chapter 2 134

Chapter 6 Characteristic Functions, Moment Generating Functions


and Related Theorems 138
6.1 Preliminaries 138
6.2 Definitions and Basic TheoremsThe One-Dimensional Case 140
Exercises 145
6.3 The Characteristic Functions of Some Random Variables 146
Exercises 149
6.4 Definitions and Basic TheoremsThe Multidimensional Case 150
Exercises 152
6.5 The Moment Generating Function and Factorial Moment Generating
Function of a Random Variable 153
Exercises 160

Chapter 7 Stochastic Independence with Some Applications 164


7.1 Stochastic Independence: Criteria of Independence 164
Exercises 168
7.2 Proof of Lemma 2 and Related Results 170
Exercises 172
7.3 Some Consequences of Independence 173
Exercises 176
7.4* Independence of Classes of Events and Related Results 177
Exercise 179

Chapter 8 Basic Limit Theorems 180


8.1 Some Modes of Convergence 180
Exercises 182
8.2 Relationships Among the Various Modes of Convergence 182
Exercises 187
8.3 The Central Limit Theorem 187
Exercises 194
8.4 Laws of Large Numbers 196
Exercises 198
x Contents

8.5 Further Limit Theorems 199


Exercises 206
8.6* Plyas Lemma and Alternative Proof of the WLLN 206
Exercises 211

Chapter 9 Transformations of Random Variables and Random


Vectors 212
9.1 The Univariate Case 212
Exercises 218
9.2 The Multivariate Case 219
Exercises 233
9.3 Linear Transformations of Random Vectors 235
Exercises 240
9.4 The Probability Integral Transform 242
Exercise 244

Chapter 10 Order Statistics and Related Theorems 245


10.1 Order Statistics and Related Distributions 245
Exercises 252
10.2 Further Distribution Theory: Probability of Coverage of a Population
Quantile 256
Exercise 258

Chapter 11 Sufficiency and Related Theorems 259


11.1 Sufficiency: Definition and Some Basic Results 260
Exercises 269
11.2 Completeness 271
Exercises 273
11.3 UnbiasednessUniqueness 274
Exercises 276
11.4 The Exponential Family of p.d.f.s: One-Dimensional Parameter
Case 276
Exercises 280
11.5 Some Multiparameter Generalizations 281
Exercises 282

Chapter 12 Point Estimation 284


12.1 Introduction 284
Exercise 284
Contents xi

12.2 Criteria for Selecting an Estimator: Unbiasedness, Minimum


Variance 285
Exercises 286
12.3 The Case of Availability of Complete Sufficient Statistics 287
Exercises 292
12.4 The Case Where Complete Sufficient Statistics Are Not Available or
May Not Exist: Cramr-Rao Inequality 293
Exercises 301
12.5 Criteria for Selecting an Estimator: The Maximum Likelihood
Principle 302
Exercises 308
12.6 Criteria for Selecting an Estimator: The Decision-
Theoretic Approach 309
12.7 Finding Bayes Estimators 312
Exercises 317
12.8 Finding Minimax Estimators 318
Exercise 320
12.9 Other Methods of Estimation 320
Exercises 322
12.10 Asymptotically Optimal Properties of Estimators 322
Exercise 325
12.11 Closing Remarks 325
Exercises 326

Chapter 13 Testing Hypotheses 327


13.1 General Concepts of the Neyman-Pearson Testing Hypotheses
Theory 327
Exercise 329
13.2 Testing a Simple Hypothesis Against a Simple Alternative 329
Exercises 336
13.3 UMP Tests for Testing Certain Composite Hypotheses 337
Exercises 347
13.4 UMPU Tests for Testing Certain Composite Hypotheses 349
Exercises 353
13.5 Testing the Parameters of a Normal Distribution 353
Exercises 356
13.6 Comparing the Parameters of Two Normal Distributions 357
Exercises 360
13.7 Likelihood Ratio Tests 361
Exercises 369
13.8 Applications of LR Tests: Contingency Tables, Goodness-of-Fit
Tests 370
Exercises 373
xii Contents

13.9 Decision-Theoretic Viewpoint of Testing Hypotheses 375

Chapter 14 Sequential Procedures 382


14.1 Some Basic Theorems of Sequential Sampling 382
Exercises 388
14.2 Sequential Probability Ratio Test 388
Exercise 392
14.3 Optimality of the SPRT-Expected Sample Size 393
Exercises 394
14.4 Some Examples 394

Chapter 15 Confidence RegionsTolerance Intervals 397


15.1 Confidence Intervals 397
Exercise 398
15.2 Some Examples 398
Exercises 404
15.3 Confidence Intervals in the Presence of Nuisance Parameters 407
Exercise 410
15.4 Confidence RegionsApproximate Confidence Intervals 410
Exercises 412
15.5 Tolerance Intervals 413

Chapter 16 The General Linear Hypothesis 416


16.1 Introduction of the Model 416
16.2 Least Square EstimatorsNormal Equations 418
16.3 Canonical Reduction of the Linear ModelEstimation of 2 424
Exercises 428
16.4 Testing Hypotheses About = E(Y) 429
Exercises 433
16.5 Derivation of the Distribution of the F Statistic 433
Exercises 436

Chapter 17 Analysis of Variance 440


17.1 One-way Layout (or One-way Classification) with the Same Number of
Observations Per Cell 440
Exercise 446
17.2 Two-way Layout (Classification) with One Observation Per Cell 446
Exercises 451
Contents xiii

17.3 2) Observations
Two-way Layout (Classification) with K (
Per Cell 452
Exercises 457
17.4 A Multicomparison method 458
Exercises 462

Chapter 18 The Multivariate Normal Distribution 463


18.1 Introduction 463
Exercises 466
18.2 Some Properties of Multivariate Normal Distributions 467
Exercise 469
18.3 Estimation of and / and a Test of Independence 469
Exercises 475

Chapter 19 Quadratic Forms 476


19.1 Introduction 476
19.2 Some Theorems on Quadratic Forms 477
Exercises 483

Chapter 20 Nonparametric Inference 485


20.1 Nonparametric Estimation 485
20.2 Nonparametric Estimation of a p.d.f. 487
Exercise 490
20.3 Some Nonparametric Tests 490
20.4 More About Nonparametric Tests: Rank Tests 493
Exercises 496
20.5 Sign Test 496
20.6 Relative Asymptotic Efficiency of Tests 498

Appendix I Topics from Vector and Matrix Algebra 499


I.1 Basic Definitions in Vector Spaces 499
I.2 Some Theorems on Vector Spaces 501
I.3 Basic Definitions About Matrices 502
I.4 Some Theorems About Matrices and Quadratic Forms 504

Appendix II Noncentral t-, 2-, and F-Distributions 508


II.1 Noncentral t-Distribution 508
xiv Contents

II.2 Noncentral x2-Distribution 508


II.3 Noncentral F-Distribution 509

Appendix III Tables 511

1 The Cumulative Binomial Distribution 511


2 The Cumulative Poisson Distribution 520
3 The Normal Distribution 523
4 Critical Values for Students t-Distribution 526
5 Critical Values for the Chi-Square Distribution 529
6 Critical Values for the F-Distribution 532
7 Table of Selected Discrete and Continuous Distributions and Some of
Their Characteristics 542

Some Notation and Abbreviations 545

Answers to Selected Exercises 547

Index 561
Contents xv

Preface to the Second Edition

This is the second edition of a book published for the first time in 1973 by
Addison-Wesley Publishing Company, Inc., under the title A First Course in
Mathematical Statistics. The first edition has been out of print for a number of
years now, although its reprint in Taiwan is still available. That issue, however,
is meant for circulation only in Taiwan.
The first issue of the book was very well received from an academic
viewpoint. I have had the pleasure of hearing colleagues telling me that the
book filled an existing gap between a plethora of textbooks of lower math-
ematical level and others of considerably higher level. A substantial number of
colleagues, holding senior academic appointments in North America and else-
where, have acknowledged to me that they made their entrance into the
wonderful world of probability and statistics through my book. I have also
heard of the book as being in a class of its own, and also as forming a collectors
item, after it went out of print. Finally, throughout the years, I have received
numerous inquiries as to the possibility of having the book reprinted. It is in
response to these comments and inquiries that I have decided to prepare a
second edition of the book.
This second edition preserves the unique character of the first issue of the
book, whereas some adjustments are affected. The changes in this issue consist
in correcting some rather minor factual errors and a considerable number of
misprints, either kindly brought to my attention by users of the book or
located by my students and myself. Also, the reissuing of the book has pro-
vided me with an excellent opportunity to incorporate certain rearrangements
of the material.
One change occurring throughout the book is the grouping of exercises of
each chapter in clusters added at the end of sections. Associating exercises
with material discussed in sections clearly makes their assignment easier. In
the process of doing this, a handful of exercises were omitted, as being too
complicated for the level of the book, and a few new ones were inserted. In

xv
xvi Contents
Preface to the Second Edition

Chapters 1 through 8, some of the materials were pulled out to form separate
sections. These sections have also been marked by an asterisk (*) to indicate
the fact that their omission does not jeopardize the flow of presentation and
understanding of the remaining material.
Specifically, in Chapter 1, the concepts of a field and of a -field, and basic
results on them, have been grouped together in Section 1.2*. They are still
readily available for those who wish to employ them to add elegance and rigor
in the discussion, but their inclusion is not indispensable. In Chapter 2, the
number of sections has been doubled from three to six. This was done by
discussing independence and product probability spaces in separate sections.
Also, the solution of the problem of the probability of matching is isolated in a
section by itself. The section on the problem of the probability of matching and
the section on product probability spaces are also marked by an asterisk for the
reason explained above. In Chapter 3, the discussion of random variables as
measurable functions and related results is carried out in a separate section,
Section 3.5*. In Chapter 4, two new sections have been created by discussing
separately marginal and conditional distribution functions and probability
density functions, and also by presenting, in Section 4.4*, the proofs of two
statements, Statements 1 and 2, formulated in Section 4.1; this last section is
also marked by an asterisk. In Chapter 5, the discussion of covariance and
correlation coefficient is carried out in a separate section; some additional
material is also presented for the purpose of further clarifying the interpreta-
tion of correlation coefficient. Also, the justification of relation (2) in Chapter 2
is done in a section by itself, Section 5.6*. In Chapter 6, the number of sections
has been expanded from three to five by discussing in separate sections charac-
teristic functions for the one-dimensional and the multidimensional case, and
also by isolating in a section by itself definitions and results on moment-
generating functions and factorial moment generating functions. In Chapter 7,
the number of sections has been doubled from two to four by presenting the
proof of Lemma 2, stated in Section 7.1, and related results in a separate
section; also, by grouping together in a section marked by an asterisk defini-
tions and results on independence. Finally, in Chapter 8, a new theorem,
Theorem 10, especially useful in estimation, has been added in Section 8.5.
Furthermore, the proof of Plyas lemma and an alternative proof of the Weak
Law of Large Numbers, based on truncation, are carried out in a separate
section, Section 8.6*, thus increasing the number of sections from five to six.
In the remaining chapters, no changes were deemed necessary, except that
in Chapter 13, the proof of Theorem 2 in Section 13.3 has been facilitated by
the formulation and proof in the same section of two lemmas, Lemma 1 and
Lemma 2. Also, in Chapter 14, the proof of Theorem 1 in Section 14.1 has been
somewhat simplified by the formulation and proof of Lemma 1 in the same
section.
Finally, a table of some commonly met distributions, along with their
means, variances and other characteristics, has been added. The value of such
a table for reference purposes is obvious, and needs no elaboration.
Preface to the Second
Contents
Edition xvii

This book contains enough material for a year course in probability and
statistics at the advanced undergraduate level, or for first-year graduate stu-
dents not having been exposed before to a serious course on the subject
matter. Some of the material can actually be omitted without disrupting the
continuity of presentation. This includes the sections marked by asterisks,
perhaps, Sections 13.413.6 in Chapter 13, and all of Chapter 14. The instruc-
tor can also be selective regarding Chapters 11 and 18. As for Chapter 19, it
has been included in the book for completeness only.
The book can also be used independently for a one-semester (or even one
quarter) course in probability alone. In such a case, one would strive to cover
the material in Chapters 1 through 10 with the exclusion, perhaps, of the
sections marked by an asterisk. One may also be selective in covering the
material in Chapter 9.
In either case, presentation of results involving characteristic functions
may be perfunctory only, with emphasis placed on moment-generating func-
tions. One should mention, however, why characteristic functions are intro-
duced in the first place, and therefore what one may be missing by not utilizing
this valuable tool.
In closing, it is to be mentioned that this author is fully aware of the fact
that the audience for a book of this level has diminished rather than increased
since the time of its first edition. He is also cognizant of the trend of having
recipes of probability and statistical results parading in textbooks, depriving
the reader of the challenge of thinking and reasoning instead delegating the
thinking to a computer. It is hoped that there is still room for a book of the
nature and scope of the one at hand. Indeed, the trend and practices just
described should make the availability of a textbook such as this one exceed-
ingly useful if not imperative.

G. G. Roussas
Davis, California
May 1996
xviii Contents

Preface to the First Edition

This book is designed for a first-year course in mathematical statistics at the


undergraduate level, as well as for first-year graduate students in statisticsor
graduate students, in generalwith no prior knowledge of statistics. A typical
three-semester course in calculus and some familiarity with linear algebra
should suffice for the understanding of most of the mathematical aspects of
this book. Some advanced calculusperhaps taken concurrentlywould be
helpful for the complete appreciation of some fine points.
There are basically two streams of textbooks on mathematical statistics
that are currently on the market. One category is the advanced level texts
which demonstrate the statistical theories in their full generality and math-
ematical rigor; for that purpose, they require a high level, mathematical back-
ground of the reader (for example, measure theory, real and complex
analysis). The other category consists of intermediate level texts, where the
concepts are demonstrated in terms of intuitive reasoning, and results are
often stated without proofs or with partial proofs that fail to satisfy an inquisi-
tive mind. Thus, readers with a modest background in mathematics and a
strong motivation to understand statistical concepts are left somewhere in
between. The advanced texts are inaccessible to them, whereas the intermedi-
ate texts deliver much less than they hope to learn in a course of mathematical
statistics. The present book attempts to bridge the gap between the two
categories, so that students without a sophisticated mathematical background
can assimilate a fairly broad spectrum of the theorems and results from math-
ematical statistics. This has been made possible by developing the fundamen-
tals of modern probability theory and the accompanying mathematical ideas at
the beginning of this book so as to prepare the reader for an understanding of
the material presented in the later chapters.
This book consists of two parts, although it is not formally so divided. Part
1 (Chapters 110) deals with probability and distribution theory, whereas Part

xviii
Contents
Preface to the First Edition xix

2 (Chapters 1120) is devoted to statistical inference. More precisely, in Part 1


the concepts of a field and -field, and also the definition of a random variable
as a measurable function, are introduced. This allows us to state and prove
fundamental results in their full generality that would otherwise be presented
vaguely using statements such as it may be shown that . . . , it can be proved
that . . . , etc. This we consider to be one of the distinctive characteristics of
this part. Other important features are as follows: a detailed and systematic
discussion of the most useful distributions along with figures and various
approximations for several of them; the establishment of several moment and
probability inequalities; the systematic employment of characteristic func-
tionsrather than moment generating functionswith all the well-known
advantages of the former over the latter; an extensive chapter on limit theo-
rems, including all common modes of convergence and their relationship; a
complete statement and proof of the Central Limit Theorem (in its classical
form); statements of the Laws of Large Numbers and several proofs of the
Weak Law of Large Numbers, and further useful limit theorems; and also an
extensive chapter on transformations of random variables with numerous
illustrative examples discussed in detail.
The second part of the book opens with an extensive chapter on suffi-
ciency. The concept of sufficiency is usually treated only in conjunction with
estimation and testing hypotheses problems. In our opinion, this does not
do justice to such an important concept as that of sufficiency. Next, the point
estimation problem is taken up and is discussed in great detail and as
large a generality as is allowed by the level of this book. Special attention is
given to estimators derived by the principles of unbiasedness, uniform mini-
mum variance and the maximum likelihood and minimax principles. An abun-
dance of examples is also found in this chapter. The following chapter is
devoted to testing hypotheses problems. Here, along with the examples (most
of them numerical) and the illustrative figures, the reader finds a discussion of
families of probability density functions which have the monotone likelihood
ratio property and, in particular, a discussion of exponential families. These
latter topics are available only in more advanced texts. Other features are
a complete formulation and treatment of the general Linear Hypothesis
and the discussion of the Analysis of Variance as an application of it.
In many textbooks of about the same level of sophistication as the present
book, the above two topics are approached either separately or in the reverse
order from the one used here, which is pedagogically unsound, although
historically logical. Finally, there are special chapters on sequential proce-
dures, confidence regionstolerance intervals, the Multivariate Normal distri-
bution, quadratic forms, and nonparametric inference.
A few of the proofs of theorems and some exercises have been drawn from
recent publications in journals.
For the convenience of the reader, the book also includes an appendix
summarizing all necessary results from vector and matrix algebra.
There are more than 120 examples and applications discussed in detail in
xx Contentsto the First Edition
Preface

the text. Also, there are more than 530 exercises, appearing at the end of the
chapters, which are of both theoretical and practical importance.
The careful selection of the material, the inclusion of a large variety of
topics, the abundance of examples, and the existence of a host of exercises of
both theoretical and applied nature will, we hope, satisfy people of both
theoretical and applied inclinations. All the application-oriented reader has to
do is to skip some fine points of some of the proofs (or some of the proofs
altogether!) when studying the book. On the other hand, the careful handling
of these same fine points should offer some satisfaction to the more math-
ematically inclined readers.
The material of this book has been presented several times to classes
of the composition mentioned earlier; that is, classes consisting of relatively
mathematically immature, eager, and adventurous sophomores, as well as
juniors and seniors, and statistically unsophisticated graduate students. These
classes met three hours a week over the academic year, and most of the
material was covered in the order in which it is presented with the occasional
exception of Chapters 14 and 20, Section 5 of Chapter 5, and Section 3 of
Chapter 9. We feel that there is enough material in this book for a three-
quarter session if the classes meet three or even four hours a week.
At various stages and times during the organization of this book several
students and colleagues helped improve it by their comments. In connection
with this, special thanks are due to G. K. Bhattacharyya. His meticulous
reading of the manuscripts resulted in many comments and suggestions that
helped improve the quality of the text. Also thanks go to B. Lind, K. G.
Mehrotra, A. Agresti, and a host of others, too many to be mentioned here. Of
course, the responsibility in this book lies with this author alone for all omis-
sions and errors which may still be found.
As the teaching of statistics becomes more widespread and its level of
sophistication and mathematical rigor (even among those with limited math-
ematical training but yet wishing to know why and how) more demanding,
we hope that this book will fill a gap and satisfy an existing need.

G. G. R.
Madison, Wisconsin
November 1972
1.1 Some Definitions and Notation 1

Chapter 1

Basic Concepts of Set Theory

1.1 Some Definitions and Notation


A set S is a (well defined) collection of distinct objects which we denote by s.
The fact that s is a member of S, an element of S, or that it belongs to S is
expressed by writing s S. The negation of the statement is expressed by
writing s S. We say that S is a subset of S, or that S is contained in S, and
write S S, if for every s S, we have s S. S is said to be a proper subset
of S, and we write S S, if S S and there exists s S such that s S. Sets
are denoted by capital letters, while lower case letters are used for elements of
sets.

S
Figure 1.1 S S; in fact, S S, since s2 S,
S'
s1 but s2 S .
s2

These concepts can be illustrated pictorially by a drawing called a Venn


diagram (Fig. 1.1). From now on a basic, or universal set, or space (which may
be different from situation to situation), to be denoted by S, will be considered
and all other sets in question will be subsets of S.

1.1.1 Set Operations


1. The complement (with respect to S) of the set A, denoted by Ac, is
defined by Ac = {s S; s A}. (See Fig. 1.2.)

1
2 1 Basic Concepts of Set Theory

S
A
Ac Figure 1.2 Ac is the shaded region.

2. The union of the sets Aj, j = 1, 2, . . . , n, to be denoted by


n
A1 A2 An or U Aj ,
j =1

is defined by
n

U A j = {s S ; s A j for at least one j = 1, 2, , n}.


j =1

For n = 2, this is pictorially illustrated in Fig. 1.3. The definition extends to an


infinite number of sets. Thus for denumerably many sets, one has

U A j = {s S ; s A j for at least one j = 1, 2, }.


j =1

Figure 1.3 A1 A2 is the shaded region.


A1 A2

3. The intersection of the sets Aj, j = 1, 2, . . . , n, to be denoted by


n
A1 A2 An or I Aj ,
j =1

is defined by
n

I A j = {s S ; s A j for all j = 1, 2, , n}.


j =1

For n = 2, this is pictorially illustrated in Fig. 1.4. This definition extends to an


infinite number of sets. Thus for denumerably many sets, one has

I A j = {s S ; s A j for all j = 1, 2, }.
j =1

Figure 1.4 A1 A2 is the shaded region.


A1 A2
1.1 Some Definitions and Notation 3

4. The difference A1 A2 is defined by

{
A1 A2 = s S ; s A1 , s A2 . }
Symmetrically,

{
A2 A1 = s S ; s A2 , s A1 . }
Note that A1 A2 = A1 Ac2, A2 A1 = A2 Ac1, and that, in general, A1 A 2
A2 A1. (See Fig. 1.5.)

S
Figure 1.5 A1 A2 is ////.
A2 A1 is \\\\.
A1 A2

5. The symmetric difference A1 A2 is defined by

( ) (
A1 A2 = A1 A2 A2 A1 . )
Note that

( ) (
A1 A2 = A1 A2 A1 A2 . )
Pictorially, this is shown in Fig. 1.6. It is worthwhile to observe that
operations (4) and (5) can be expressed in terms of operations (1), (2), and
(3).

Figure 1.6 A1 A2 is the shaded area.


A1 A2

1.1.2 Further Definitions and Notation


A set which contains no elements is called the empty set and is denoted by .
Two sets A1, A2 are said to be disjoint if A1 A2 = . Two sets A1, A2 are said
to be equal, and we write A1 = A2, if both A1 A2 and A2 A1. The sets Aj,
j = 1, 2, . . . are said to be pairwise or mutually disjoint if Ai Aj = for all
i j (Fig. 1.7). In such a case, it is customary to write
n
A1 + A2 , A1 + + An = A j and A1 + A2 + = A j
j =1 j =1

instead of A1 A2, U A j, and U A j, respectively. We will write
n
j =1 j =1
U j A j, j A j, I j A j, where we do not wish to specify the range of j, which
4 1 Basic Concepts of Set Theory

will usually be either the (finite) set {1, 2, . . . , n}, or the (infinite) set
{1, 2, . . .}.

S
Figure 1.7 A1 and A2 are disjoint; that is,
A1 A2 = . Also A1 A2 = A1 + A2 for the
same reason.
A1 A2

1.1.3 Properties of the Operations on Sets


1. S c = , c = S, (Ac)c = A.
2. S A = S, A = A, A Ac = S, A A = A.
3. S A = A, A = , A Ac = , A A = A.
The previous statements are all obvious as is the following: A for every
subset A of S. Also
4. A1 (A2 A3) = (A1 A2) A3
A1 (A2 A3) = (A1 A2) A3 } (Associative laws)

5. A1 A2 = A2 A1
A1 A2 = A2 A1 } (Commutative laws)

6. A (j Aj) = j (A Aj)
A (j Aj) = j (A Aj) } (Distributive laws)

are easily seen to be true.


The following identity is a useful tool in writing a union of sets as a sum of
disjoint sets.
An identity:

UA j = A1 + A1c A2 + A1c A2c A3 + .


j

There are two more important properties of the operation on sets which
relate complementation to union and intersection. They are known as De
Morgans laws:
c

i ) U Aj = I Aj ,
j j
c

c

ii ) I Aj = U Aj .
j j
c

As an example of a set theoretic proof, we prove (i).


PROOF OF (i) We wish to establish

a) ( U jAj)c I jAcj and b) I jAcj ( U j Aj)c.


1.1 Some Definitions andExercises
Notation 5

We will then, by definition, have verified the desired equality of the two
sets.
a) Let s ( U jAj)c. Then s U jAj, hence s Aj for any j. Thus s Acj for every
j and therefore s I jAcj.
b) Let s I jAcj. Then s Acj for every j and hence s Aj for any j. Then
s U jAj and therefore s ( U jAj)c.
The proof of (ii) is quite similar.

This section is concluded with the following:


DEFINITION 1 The sequence {An}, n = 1, 2, . . . , is said to be a monotone sequence of sets if:
ii) A1 A2 A3 (that is, An is increasing, to be denoted by An), or
ii) A1 ! A2 ! A3 ! (that is, An is decreasing, to be denoted by An).
The limit of a monotone sequence is defined as follows:

ii) If An, then lim An = U An , and
n
n=1


ii) If An, then lim An = I An .
n
n=1

More generally, for any sequence {An}, n = 1, 2, . . . , we define



A = lim inf An = U I Aj,
n
n=1 j= n

and

A = lim sup An = I U Aj .
n n=1 j= n

The sets A and A are called the inferior limit and superior limit,
.
respectively, of the sequence {An}. The sequence {An} has a limit if A = A

Exercises
1.1.1 Let Aj, j = 1, 2, 3 be arbitrary subsets of S. Determine whether each of
the following statements is correct or incorrect.
iii) (A1 A2) A2 = A2;
iii) (A1 A2) A1 = A2;
iii) (A1 A2) (A1 A2) = ;
iv) (A1 A2) (A2 A3) (A3 A1) = (A1 A2) (A2 A3) (A3 A1).
6 1 Basic Concepts of Set Theory

1.1.2 Let S = {(x, y) ! 2; 5 ! x ! 5, 0 ! y ! 5, x, y = integers}, where


prime denotes transpose, and define the subsets Aj, j = 1, . . . , 7 of S as follows:



( )
A1 = x, y S ; x = y;

( ) S ; x = y;
A2 = x, y



( )
A3 = x, y S ; x 2 = y 2 ;

( ) S ; x
A4 = x, y

2
y 2 ;



( )
A5 = x, y S ; x 2 + y 2 4 ;

( )
A6 = x, y S ; x y2 ;




( )
A7 = x, y S ; x 2 y.

List the members of the sets just defined.


1.1.3 Refer to Exercise 1.1.2 and show that:

7 7
(
iii) A1 U A j = U A1 A j ;
j =2 j =2
)
7 7
(
iii) A1 I A j = I A1 A j ;
j =2 j =2
)
c
7 7
iii) U A j = I Acj ;
j =1 j =1

c
7 7
iv) I A j = U Acj
j =1 j =1

by listing the members of each one of the eight sets appearing on either side of
each one of the relations (i)(iv).

1.1.4 Let A, B and C be subsets of S and suppose that A B and B C.


Then show that A C; that is, the subset relationship is transitive. Verify it by
taking A = A1, B = A3 and C = A4, where A1,A3 and A4 are defined in Exercise
1.1.2.

1.1.5 Establish the distributive laws stated on page 4.


1.1.6 In terms of the acts A1, A2, A3, and perhaps their complements,
express each one of the following acts:

iii) Bi = {s S; s belongs to exactly i of A1, A2, A3 , where i = 0, 1, 2, 3};


iii) C = {s S; s belongs to all of A1, A2, A3 };
1.1 Some Definitions andExercises
Notation 7

iii) D = {s S; s belongs to none of A1, A2, A3};


iv) E = {s S; s belongs to at most 2 of A1, A2, A3};
iv) F = {s S; s belongs to at least 1 of A1, A2, A3 }.

1.1.7 Establish the identity stated on page 4.


1.1.8 Give a detailed proof of the second identity in De Morgans laws; that
is, show that

c

I Aj = U Aj .
c

j j

1.1.9 Refer to Definition 1 and show that


iii) A = {s S; s belongs to all but finitely many As};
iii) A = {s S; s belongs to infinitely many As};
;
iii) A A

iv) If {An} is a monotone sequence, then A = A = lim An .
n

1.1.10 Let S = ! 2 and define the subsets An, Bn, n = 1, 2, . . . of S as


follows:

1
( ) 1 2
An = x, y ! 2 ; 3 + x < 6 , 0 y 2 2 ,
n n n

1
( )
Bn = x, y ! 2 ; x 2 + y 2 3 .
n

Then show that An A, Bn B and identify A and B.


1.1.11 Let S = ! and define the subsets An, Bn, n = 1, 2, . . . of S as
follows:

1 1 3
An = x ! ; 5 + < x < 20 , Bn = x ! ; 0 < x < 7 + .
n n n

Then show that An and Bn, so that lim n


An = A and lim B = B exist (by
n n
Exercise 1.1.9(iv)). Also identify the sets A and B.
1.1.12 Let A and B be subsets of S and for n = 1, 2, . . . , define the sets An as
follows: A2n1 = A, A2n = B. Then show that

lim inf An = A B, lim sup An = A B.


n n
8 1 Basic Concepts of Set Theory

1.2* Fields and -Fields


In this section, we introduce the concepts of a field and of a -field, present a
number of examples, and derive some basic results.
DEFINITION 2 A class (set) of subsets of S is said to be a field, and is denoted by F, if
(F1) F is a non-empty class.
(F2) A F implies that Ac F (that is, F is closed under
complementation).
(F3) A1, A2 F implies that A1 A2 F (that is, F is closed under
pairwise unions).

1.2.1 Consequences of the Definition of a Field


1. S, F.
2. If Aj F, j = 1, 2, . . . , n, then Unj =1 A j F, Inj =1 A j F for any finite n.
(That is, F is closed under finite unions and intersections. Notice, how-
ever, that Aj F, j = 1, 2, . . . need not imply that their union or intersection is
in F; for a counterexample, see consequence 2 on page 10.)
PROOF OF (1) AND (2) (1) (F1) implies that there exists A F and (F2)
implies that Ac F. By (F3), A Ac = S F. By (F2), S c = F.
(2) The proof will be by induction on n and by one of the De Morgans
laws. By (F3), if A1, A2 F, then A1 A2 F; hence the statement for unions
is true for n = 2. (It is trivially true for n = 1.) Now assume the statement for
unions is true for n = k 1; that is, if
k1
A1 , A2 , , Ak1 F , then U Aj F .
j =1

Consider A1, A2, . . . , Ak F. By the associative law for unions of sets,


k k1
U j U A j Ak .
A =
j =1 j =1

By the induction hypothesis, Ukj =11 A j F. Since Ak F, (F3) implies that

k1 k

U A j Ak = U A j F
j =1 j =1

and by induction, the statement for unions is true for any finite n. By observing
that
c
n n c
=
I j U A j ,
A
j =1 j =1

* The reader is reminded that sections marked by an asterisk may be omitted without jeo-
* pardizing the understanding of the remaining material.
1.1 Some1.2*
Definitions
Fields and -Fields
and Notation 9

we see that (F2) and the above statement for unions imply that if A1, . . . , An
F, then Inj =1 A j F for any finite n.

1.2.2 Examples of Fields


1. C1 = {, S} is a field (trivial field).
2. C2 = {all subsets of S} is a field (discrete field).
3. C3 = {, S, A, Ac}, for some A S, is a field.
4. Let S be infinite (countably so or not) and let C4 be the class of subsets
of S which are finite, or whose complements are finite; that is, C4 = {A S; A
or Ac is finite}.
As an example, we shall verify that C4 is a field.
PROOF THAT C4 IS A FIELD
i) Since S c = is finite, S C4, so that C4 is non-empty.
ii) Suppose that A C4. Then A or Ac is finite. If A is finite, then (Ac)c = A is
finite and hence Ac C4 also. If Ac is finite, then Ac C4.
iii) Suppose that A1, A2 C4. Then A1 or Ac1 is finite and A2 or Ac2 is finite.

a) Suppose that A1, A2 are both finite. Then A1 A2 is finite, so that A1


A2 C4.
b) Suppose that A1c, A2 are finite. Then (A1 A2)c = A1c Ac2 is finite
since A1c is. Hence A1 A2 C4.
The other two possibilities follow just as in (b). Hence (F1), (F2), (F3) are
satisfied.
We now formulate and prove the following theorems about fields.
THEOREM 1 Let I be any non-empty index set (finite, or countably infinite, or uncoun-
table), and let Fj, j I be fields of subsets of S. Define F by F = I j I F j =
{A; A Fj for all j I}. Then F is a field.
PROOF
i) S, Fj for every j I, so that S, F and hence F is non-empty.
ii) If A F, then A Fj for every j I. Thus Ac Fj for every j I, so that
Ac F.
iii) If A1, A2 F, then A1, A2 Fj for every j I. Then A1 A2 Fj for every
j I, and hence A1 A2 F.
THEOREM 2 Let C be an arbitrary class of subsets of S. Then there is a unique minimal field
F containing C. (We say that F is generated by C and write F = F(C).)
PROOF Clearly, C is contained in the discrete field. Next, let {Fj, j I} be the
class of all fields containing C and define F(C) by
( )
F C = I F j.
j I
10 1 Basic Concepts of Set Theory

By Theorem 1, F(C) is a field containing C. It is obviously the smallest such


field, since it is the intersection of all fields containing C, and is unique. Hence
F = F(C).
DEFINITION 3 A class of subsets of S is said to be a -field, and is denoted by A, if it is a field
and furthermore (F3) is replaced by (A3): If Aj A, j = 1, 2, . . . , then U j = 1 A j
A (that is, A is closed under denumerable unions).

1.2.3 Consequences of the Definition of a -Field


1. If Aj A, j = 1, 2, . . . , then I j = 1 A j A (that is, A is closed under
denumerable intersections).
2. By definition, a -field is a field, but the converse is not true. In fact, in
Example 4 on page 9, take S = (, ), and define Aj = {all integers in [j, j]},
j = 1, 2, . . . . Then U j = 1 A j is the set A, say, of all integers. Thus A is infinite and
furthermore so is Ac. Hence A F, whereas Aj F for all j.

1.2.4 Examples of -Fields


1. C1 = {, S} is a -field (trivial -field).
2. C2 = {all subsets of S} is a -field (discrete -field).
3. C3 = {, S, A, Ac} for some A S is a -field.
4. Take S to be uncountable and define C4 as follows:
C4 = {all subsets of S which are countable or whose complements are
countable}.
As an example, we prove that C4 is a -field.
PROOF
i) Sc = is countable, so C4 is non-empty.
ii) If A C4, then A or Ac is countable. If A is countable, then (Ac)c = A is
countable, so that Ac C4. If Ac is countable, by definition Ac C4.
iii) The proof of this statement requires knowledge of the fact that a count-
able union of countable sets is countable. (For proof of this fact see
page 36 in Tom M. Apostols book Mathematical Analysis, published
by Addison-Wesley, 1957.) Let Aj, j = 1, 2, . . . A. Then either each
Aj is countable, or there exists some Aj for which Aj is not countable but
Acj is. In the first case, we invoke the previously mentioned theorem
on the countable union of countable sets. In the second case, we note
that
c

U j A = I Acj ,
j = 1 j = 1
which is countable, since it is the intersection of sets, one of which is
countable.
1.1 Some1.2*
Definitions
Fields and -Fields
and Notation 11

We now introduce some useful theorems about -fields.


THEOREM 3 Let I be as in Theorem 1, and let Aj, j I, be -fields. Define A by A =
I jI Aj = {A; A Aj for all j I}. Then A is a -field.
PROOF
i) S, Aj for every j I and hence they belong in A.
ii) If A A, then A Aj for every j I, so that Ac Aj for every j I. Thus
Ac A.
iii) If A1, A2, . . . , A, then A1, A2, . . . Aj for every j I and hence U j = 1 A j
A j; for every j I; thus U j = 1 A j A .

THEOREM 4 Let C be an arbitrary class of subsets of S. Then there is a unique minimal


-field A containing C. (We say that A is the -field generated by C and write
A = (C).)

PROOF Clearly, C is contained in the discrete -field. Define

( ) {
C = I all -fields containing C . }
By Theorem 3, (C ) is a -field which obviously contains C. Uniqueness
and minimality again follow from the definition of (C ). Hence A =
(C).

REMARK 1 For later use, we note that if A is a -field and A A, then AA


= {C; C = B A for some B A} is a -field, where complements of sets are
formed with respect to A, which now plays the role of the entire space. This is
easily seen to be true by the distributive property of intersection over union
(see also Exercise 1.2.5).
In all that follows, if S is countable (that is, finite or denumerably in-
finite), we will always take A to be the discrete -field. However, if S is
uncountable, then for certain technical reasons, we take the -field to be
smaller than the discrete one. In both cases, the pair (S, A) is called a
measurable space.

1.2.5 Special Cases of Measurable Spaces


1. Let S be ! (the set of real numbers, or otherwise known as the real
line) and define C0 as follows:

( )( ]( )[
, x , , x , x, , x, , x, y ,
)(
)
{ }
C0 = all intervals in ! = .
( ][ )[ ]
x, y , x, y , x, y ; x, y ! , x < y

By Theorem 4, there is a -field A = (C0); we denote this -field by B and call


12 1 Basic Concepts of Set Theory

it the Borel -field (over the real line). The pair (! , B) is called the Borel real
line.
THEOREM 5 Each one of the following classes generates the Borel -field.
C1 = {(x, y]; x, y !, x < y},
C2 = {[ x, y); x, y ! , x < y},

C3 = {[ x, y]; x, y ! , x < y},

C4 = {( x, y); x, y ! , x < y},

C5 = {( x, ); x !},

C6 = {[ x, ); x !},

C7 = {( , x ); x !},

C8 = {( , x ]; x !}.
Also the classes C j, j = 1, . . . , 8 generate the Borel -field, where for j = 1, . . . ,
8, Cj is defined the same way as Cj is except that x, y are restricted to the
rational numbers.

PROOF Clearly, if C, C are two classes of subsets of S such that C C, then


(C) (C). Thus, in order to prove the theorem, it suffices to prove that B
(Cj), B (Cj), j = 1, 2, . . . , 8, and in order to prove this, it suffices to show
that C0 (Cj), C0 (Cj), j = 1, 2, . . . , 8. As an example, we show that C0
(C7). Consider xn x. Then (, xn) (C7) and hence In=1 (, xn) (C7).
But

I (, x ) = (, x ].
n
n=1

Thus (, x] (C7) for every x ! . Since

(x, ) = (, x] , [x, ) = (, x) ,
c c

it also follows that (x, ), [x, ) (C7). Next,

(x, y) = (, y) (, x] = (, y) (x, ) (C ), 7

(x, y] = (, y] (x, ) (C ), [x, y) = (, y) [x, ) (C ),


7 7

[x, y] = (, y] [x, ) (C ). 7

Thus C0 (C7). In the case of Cj, j = 1, 2, . . . , 8, consider monotone


sequences of rational numbers convergent to given irrationals x, y.
1.1 Some Definitions andExercises
Notation 13

2. Let S = ! ! = ! 2 and define C0 as follows:

{ } {(, x) (, x ), (, x) (, x ],
C0 = all rectangles in ! 2 =

(, x] (, x ), (, x] (, x ],
(x, ) (x , ), , [x, ) [x , ), ,
(x, y) (x , y), , [x, y] [x , y],
x, y, x , y ! , x < y, x < y}.

The -field generated by C0 is denoted by B2 and is the two-dimensional


Borel -field. A theorem similar to Theorem 5 holds here too.
3. Let S = ! ! ! = ! k (k copies of ! ) and define C0 in a way
similar to that in (2) above. The -field generated by C0 is denoted by Bk and
is the k-dimensional Borel -field. A theorem similar to Theorem 5 holds here
too.

Exercises
1.2.1 Verify that the classes defined in Examples 1, 2 and 3 on page 9 are
fields.
1.2.2 Show that in the definition of a field (Definition 2), property (F3) can
be replaced by (F3) which states that if A1, A2 F, then A1 A2 F.
1.2.3 Show that in Definition 3, property (A3) can be replaced by (A3),
which states that if

Aj A, j = 1, 2, then I Aj A.
j =1

1.2.4 Refer to Example 4 on -fields on page 10 and explain why S was taken
to be uncountable.
1.2.5 Give a formal proof of the fact that the class AA defined in Remark 1
is a -field.
1.2.6 Refer to Definition 1 and show that all three sets A, A and lim An,
n
whenever it exists, belong to A provided An, n 1, belong to A.
1.2.7 Let S = {1, 2, 3, 4} and define the class C of subsets of S as follows:

{ {} {} {} { } { } { } { } { } { }
C = , 1 , 2 , 3 , 4 , 1, 2 , 1, 3 , 1, 4 , 2, 3 , 2, 4 ,

{1, 2, 3}, {1, 3, 4}, {2, 3, 4}, S }.


Determine whether or not C is a field.
1.2.8 Complete the proof of the remaining parts in Theorem 5.
14 2 Some Probabilistic Concepts and Results

Chapter 2

Some Probabilistic Concepts


and Results

2.1 Probability Functions and Some Basic Properties and Results


Intuitively by an experiment one pictures a procedure being carried out under
a certain set of conditions whereby the procedure can be repeated any number
of times under the same set of conditions, and upon completion of the proce-
dure certain results are observed. An experiment is a deterministic experiment
if, given the conditions under which the experiment is carried out, the outcome
is completely determined. If, for example, a container of pure water is brought
to a temperature of 100C and 760 mmHg of atmospheric pressure the out-
come is that the water will boil. Also, a certificate of deposit of $1,000 at the
annual rate of 5% will yield $1,050 after one year, and $(1.05)n 1,000 after n
years when the (constant) interest rate is compounded. An experiment for
which the outcome cannot be determined, except that it is known to be one of
a set of possible outcomes, is called a random experiment. Only random
experiments will be considered in this book. Examples of random experiments
are tossing a coin, rolling a die, drawing a card from a standard deck of playing
cards, recording the number of telephone calls which arrive at a telephone
exchange within a specified period of time, counting the number of defective
items produced by a certain manufacturing process within a certain period of
time, recording the heights of individuals in a certain class, etc. The set of all
possible outcomes of a random experiment is called a sample space and is
denoted by S. The elements s of S are called sample points. Certain subsets of
S are called events. Events of the form {s} are called simple events, while an
event containing at least two sample points is called a composite event. S and
are always events, and are called the sure or certain event and the impossible
event, respectively. The class of all events has got to be sufficiently rich in order
to be meaningful. Accordingly, we require that, if A is an event, then so is its
complement Ac. Also, if Aj, j = 1, 2, . . . are events, then so is their union U j Aj.

14
2.1 Probability Functions and Some Basic
2.4 Properties
Combinatorial
and Results 15

(In the terminology of Section 1.2, we require that the events associated with
a sample space form a -field of subsets in that space.) It follows then that I j Aj
is also an event, and so is A1 A2, etc. If the random experiment results in s and
s A, we say that the event A occurs or happens. The U j Aj occurs if at least
one of the Aj occurs, the I j Aj occurs if all Aj occur, A1 A2 occurs if A1 occurs
but A2 does not, etc.
The next basic quantity to be introduced here is that of a probability
function (or of a probability measure).
DEFINITION 1 A probability function denoted by P is a (set) function which assigns to each
event A a number denoted by P(A), called the probability of A, and satisfies
the following requirements:
(P1) P is non-negative; that is, P(A) 0, for every event A.
(P2) P is normed; that is, P(S) = 1.
(P3) P is -additive; that is, for every collection of pairwise (or mutually)
disjoint events Aj, j = 1, 2, . . . , we have P(j Aj) = j P(Aj).
This is the axiomatic (Kolmogorov) definition of probability. The triple
(S, class of events, P) (or (S, A, P)) is known as a probability space.
REMARK 1 If S is finite, then every subset of S is an event (that is, A is taken
to be the discrete -field). In such a case, there are only finitely many events
and hence, in particular, finitely many pairwise disjoint events. Then (P3) is
reduced to:
(P3) P is finitely additive; that is, for every collection of pairwise disjoint
events, Aj, j = 1, 2, . . . , n, we have
n n

( )
P A j = P A j .
j =1 j =1
Actually, in such a case it is sufficient to assume that (P3) holds for any two
disjoint events; (P3) follows then from this assumption by induction.

2.1.1 Consequences of Definition 1


(C1) P() = 0. In fact, S = S + + ,
so that

( ) ( ) ( ) ( )
P S = P S ++ = P S +P + ,
or

( )
1 = 1+ P + and ( )
P = 0,
since P() 0. (So P() = 0. Any event, possibly " , with probability 0 is
called a null event.)
(C2) P is finitely additive; that is for any event Aj, j = 1, 2, . . . , n such that
Ai Aj = , i j,
16 2 Some Probabilistic Concepts and Results

n n
P A j = P A j .
j =1 j =1
( )
Indeed, for ( ) ( ) ( )
Aj = 0, j n + 1, P jn= 1Aj = P j= 1PAj = j= 1P Aj =
( )
jn= 1P Aj .
(C3) For every event A, P(Ac) = 1 P(A). In fact, since A + Ac = S,

( ) ()
P A + Ac = P S , or ( ) ( )
P A + P Ac = 1,

so that P(Ac) = 1 P(A).


(C4) P is a non-decreasing function; that is A1 A2 implies P(A1) P(A2).
In fact,

(
A2 = A1 + A2 A1 , )
hence

( ) ( ) (
P A2 = P A1 + P A2 A1 , )
and therefore P(A2) P(A1).
REMARK 2 If A1 A2, then P(A2 A1) = P(A2) P(A1), but this is not true,
in general.
(C5) 0 P(A) 1 for every event A. This follows from (C1), (P2), and (C4).
(C6) For any events A1, A2, P(A1 A2) = P(A1) + P(A2) P(A1 A2).
In fact,

(
A1 A2 = A1 + A2 A1 A2 . )
Hence

( ) ( ) (
P A1 A2 = P A1 + P A2 A1 A2 )
= P( A ) + P( A ) P( A A ),
1 2 1 2

since A1 A2 A2 implies

( ) ( ) (
P A2 A1 A2 = P A2 P A1 A2 . )
(C7) P is subadditive; that is,


P U A j P A j
j =1 j =1
( )
and also
n n
P U A j P A j .
j =1 j =1
( )
2.1 Probability Functions and Some Basic
2.4 Properties
Combinatorial
and Results 17

This follows from the identities

U A j = A1 + ( A1c A2 ) + + ( A1c Anc1 An ) + ,


j =1

U A j = A1 + ( A1c A2 ) + + ( A1c Anc1 An ),


n

j =1

(P3) and (C2), respectively, and (C4).


A special case of a probability space is the following: Let S = {s1, s2, . . . , sn},
let the class of events be the class of all subsets of S, and define P as P({sj}) =
1/n, j = 1, 2, . . . , n. With this definition, P clearly satisfies (P1)(P3) and this
is the classical definition of probability. Such a probability function is called a
uniform probability function. This definition is adequate as long as S is finite
and the simple events {sj}, j = 1, 2, . . . , n, may be assumed to be equally
likely, but it breaks down if either of these two conditions is not satisfied.
However, this classical definition together with the following relative frequency
(or statistical) definition of probability served as a motivation for arriving at
the axioms (P1)(P3) in the Kolmogorov definition of probability. The relative
frequency definition of probability is this: Let S be any sample space, finite
or not, supplied with a class of events A. A random experiment associated with
the sample space S is carried out n times. Let n(A) be the number of times that
the event A occurs. If, as n , lim[n(A)/n] exists, it is called the probability
of A, and is denoted by P(A). Clearly, this definition satisfies (P1), (P2) and
(P3).
Neither the classical definition nor the relative frequency definition of
probability is adequate for a deep study of probability theory. The relative
frequency definition of probability provides, however, an intuitively satisfac-
tory interpretation of the concept of probability.
We now state and prove some general theorems about probability
functions.

THEOREM 1 (Additive Theorem) For any finite number of events, we have

n
( )
n

( )
P U A j = P A j P A j A j
j =1 j =1 1 j < j n
1 2
1 2

+
1 j1 < j2 < j3 n
(
P Aj Aj Aj
1 2 3
)
( ) ( )
n+1
+ 1 P A1 A2 An .

PROOF (By induction on n). For n = 1, the statement is trivial, and we


have proven the case n = 2 as consequence (C6) of the definition of probability
functions. Now assume the result to be true for n = k, and prove it for
n = k + 1.
18 2 Some Probabilistic Concepts and Results

We have
k+1 k
P U A j = P U A j Ak+1
j =1 j =1
k k
j =1 j =1
(
= P U A j + P Ak+1 P U A j Ak+1

)
k
( )
= P A j P A j1 A j2
j = 1 1 j1 < j2 k
( )
+ P A j1 A j2 A j3
1 j1 < j2 < j3 k
( )
k
( ) (
P A1 A2 Ak + P Ak+1 P U A j Ak+1 ) ( ) ( )
k+1
+ 1
j =1
k+1
= P Aj
j =1
( ) P( A A ) 1 ji < j2 k
j1 j2

+ P( A A A ) j1 j2 j3
1 ji < j2 < j3 k

k
( ) ( ) ( ) .
k+1
+ 1 P A1 A2 Ak P U A j Ak+1 (1)
But j =1

k
( )
k

( )
P U A j Ak+1 = P A j Ak+1 P A j1 A j2 Ak+1
j =1 j =1 1 j1 < j2 k
( )
+
1 j1 < j2 < j3 k
(
P A j1 A j2 A j3 Ak+1 )
( ) ( )
k
+ 1 P A j1 A jk 11 Ak+1
1 j1 < j2 jk 1 k

( ) ( )
k+1
+ 1 P A1 Ak Ak+1 .

Replacing this in (1), we get

k+1 k+1
( )
k

( )
P U A j = P A j P A j A j + P A j Ak+1
j =1 j =1 1 j < j k j =1
1 2
( )
1 2


(
+ P A j A j A j + P A j A j Ak+1
1 j < j < j k 1 j < j k
1 2 3
) ( 1 2
)
1 2 3 1 2

( ) [ P( A A )
k+1
+ 1 1 k


+
1 j1 < j2 < < jk 1 k
(
P A j A j Ak+1
1

k 1
)
2.1 Probability Functions and Some Basic
2.4 Properties
Combinatorial
and Results 19

( ) P( A A A )
k+ 2
+ 1 1 k k+1
k+1
= P( A ) P( A A )
j j1 j2
j=1 1 j1 < j2 k+1

+ P( A A A ) j1 j2 j3
1 j1 < j2 < j3 k+1

( ) ( )
k+ 2
+ 1 P A1 Ak+1 .

THEOREM 2 Let {An} be a sequence of events such that, as n , An or An . Then,

(
P lim An = lim P An .
n
) n
( )
PROOF Let us first assume that An . Then

lim An = U A j.
n
j=1

We recall that

( ) ( )

UA j = A1 + A1c A2 + A1c A2c A3 +


j =1

(
= A1 + A2 A1 + A3 A2 + , ) ( )
by the assumption that An . Hence

(
)
P lim An = P U A j = P A1 + P A2 A1
n j =1
( ) ( )
( ( ) )
+ P A3 A2 + + P An An1 +
= lim[ P( A ) + P( A A ) + + P( A A )]
1 2 1 n n1
n

= lim[ P( A ) + P( A ) P( A )
1 2 1
n

+ P( A ) P( A ) + + P( A ) P( A )]
3 2 n n1

= lim P( A ). n
n

Thus

(
P lim An = lim P An .
n
) n
( )
Now let An . Then Acn , so that

lim Anc = U Acj .
n
j=1
20 2 Some Probabilistic Concepts and Results

Hence

(
n
)

j=1 n
( )
P lim Anc = P U Acj = lim P Anc ,

or equivalently,

c


j = 1 n [ ( )]
P I A j = lim 1 P An , or 1 P I A j = 1 lim P An .
j =1 n
( )

Thus

n

( )

(
lim P An = P I A j = P lim An ,
j =1 n
)
and the theorem is established.
This theorem will prove very useful in many parts of this book.

Exercises
2.1.1 If the events Aj, j = 1, 2, 3 are such that A1 A2 A3 and P(A1) = 1
4
,
P(A2) = 125 , P(A3) = 7 , compute the probability of the following events:
12

A A2 , A1c A3 , A2c A3 , A1 A2c A3c , A1c A2c A3c .


1
c

2.1.2 If two fair dice are rolled once, what is the probability that the total
number of spots shown is
i) Equal to 5?
ii) Divisible by 3?
2.1.3 Twenty balls numbered from 1 to 20 are mixed in an urn and two balls
are drawn successively and without replacement. If x1 and x2 are the numbers
written on the first and second ball drawn, respectively, what is the probability
that:
i) x1 + x2 = 8?
ii) x1 + x2 5?
2.1.4 Let S = {x integer; 1 x 200} and define the events A, B, and C by:

{
A = x S ; x is divisible by 7 , }
B = {x S ; x = 3n + 10 }
for some positive integer n ,

{
C = x S ; x 2 + 1 375 . }
2.2
2.4 Conditional
Combinatorial
Probability
Results 21

Compute P(A), P(B), P(C), where P is the equally likely probability function
on the events of S.
2.1.5 Let S be the set of all outcomes when flipping a fair coin four times and
let P be the uniform probability function on the events of S. Define the events
A, B as follows:

{ }
A = s S ; s contains more T s than H s ,
B = {s S ; any T in s precedes every H in s}.

Compute the probabilities P(A), P(B).


2.1.6 Suppose that the events Aj, j = 1, 2, . . . are such that

P( A j ) < .
j=1

Use Definition 1 in Chapter 1 and Theorem 2 in this chapter in order to show


) = 0.
that P(A
2.1.7 Consider the events Aj, j = 1, 2, . . . and use Definition 1 in Chapter 1
and Theorem 2 herein in order to show that

( ) ( ) ( )
P A lim inf P An lim sup P An P A .
n n
( )

2.2 Conditional Probability


In this section, we shall introduce the concepts of conditional probability and
stochastic independence. Before the formal definition of conditional probabil-
ity is given, we shall attempt to provide some intuitive motivation for it. To this
end, consider a balanced die and suppose that the sides bearing the numbers
1, 4 and 6 are painted red, whereas the remaining three sides are painted black.
The die is rolled once and we are asked for the probability that the upward
side is the one bearing the number 6. Assuming the uniform probability
function, the answer is, clearly, 16 . Next, suppose that the die is rolled once as
before and all that we can observe is the color of the upward side but not the
number on it (for example, we may be observing the die from a considerable
distance, so that the color is visible but not the numbers on the sides). The
same question as above is asked, namely, what is the probability that the
number on the uppermost side is 6. Again, by assuming the uniform probabil-
ity function, the answer now is 13 . This latter probability is called the condi-
tional probability of the number 6 turning up, given the information that the
uppermost side was painted red. Letting B stand for the event that number 6
appears and A for the event that the uppermost side is red, the above-
22 2 Some Probabilistic Concepts and Results

mentioned conditional probability is denoted by P(B|A), and we observe that


this is equal to the quotient P(A B)/P(A). Or suppose that, for the purposes
of a certain study, we observe two-children families in a certain locality, and
record the gender of the children. A sample space for this experiment is the
following: S = {bb, bg, gb, gg}, where b stands for boy and g for girl, and bg, for
example, indicates that the boy is older than the girl. Suppose further
(although this is not exactly correct) that: P({bb}) = P({bg}) = P({gb}) = P({gg})
= 14 , and define the events A and B as follows: A = children of one gender =
{bb, gg}, B = at least one boy = {bb, bg, gb}. Then P(A|B) = P(A B)/
P(B) = 13 .
From these and other examples, one is led to the following definition of
conditional probability.

DEFINITION 2 Let A be an event such that P(A) > 0. Then the conditional probability, given
A, is the (set) function denoted by P(|A) and defined for every event B as
follows:

( ) (P( A) ) .
P A B
P BA =

P(B|A) is called the conditional probability of B, given A.


The set function P(|A) is actually a probability function. To see this, it
suffices to prove the P(|A) satisfies (P1)(P3). We have: P(B|A) # 0 for every
event B, clearly. Next,

( ) (P( A) ) = P(( A)) = 1,


P SA P A
PSA =

and if Aj, j = 1, 2, . . . , are events such that Ai Aj = , i j, we have



(
P j = 1 A j A P j = 1 A j A

)
P A j A = =
j =1 P A ( ) P A ( )
(
P Aj A )=
P( A )

P( A )
1
= A = A.
( )
P A j =1
j
j =1 ( )
P A j =1
j

The conditional probability can be used in expressing the probability of


the intersection of a finite number of events.

THEOREM 3 (Multiplicative Theorem) Let Aj, j = 1, 2, . . . , n, be events such that

n1
P I A j > 0.
j =1
Then
2.4
2.2 Conditional
Combinatorial
Probability
Results 23

n
(
P I A j = P An A1 A2 An1
j =1
)
(
P An1 A1 An 2 P A2 A1 P A1 . ) ( )( )
(The proof of this theorem is left as an exercise; see Exercise 2.2.4.)
REMARK 3 The value of the above formula lies in the fact that, in general, it
is easier to calculate the conditional probabilities on the right-hand side. This
point is illustrated by the following simple example.
EXAMPLE 1 An urn contains 10 identical balls (except for color) of which five are black,
three are red and two are white. Four balls are drawn without replacement.
Find the probability that the first ball is black, the second red, the third white
and the fourth black.
Let A1 be the event that the first ball is black, A2 be the event that the
second ball is red, A3 be the event that the third ball is white and A4 be the
event that the fourth ball is black. Then

(
P A1 A2 A3 A4 )
( )(
= P A4 A1 A2 A3 P A3 A1 A2 P A2 A1 P A1 , )( )( )
and by using the uniform probability function, we have

( )
P A1 =
5
10
, (
P A2 A1 = ) 3
9
, (
P A3 A1 A2 = ) 2
8
,

(
P A4 A1 A2 A3 = ) 4
7
.

Thus the required probability is equal to 1


42
! 0.0238.
Now let Aj, j = 1, 2, . . . , be events such that Ai Aj = , i j, and j Aj =
S. Such a collection of events is called a partition of S. The partition is finite
or (denumerably) infinite, accordingly, as the events Aj are finitely or
(denumerably) infinitely many. For any event, we clearly have:

(
B = B Aj .
j
)
Hence

( ) (
P B = P B Aj = P B Aj P Aj ,
j
) j
( )( )
provided P(Aj) > 0, for all j. Thus we have the following theorem.
THEOREM 4 (Total Probability Theorem) Let {Aj, j = 1, 2, . . . } be a partition of S with
P(Aj) > 0, all j. Then for B A, we have P(B) = jP(B|Aj)P(Aj).
This formula gives a way of evaluating P(B) in terms of P(B|Aj) and
P(Aj), j = 1, 2, . . . . Under the condition that P(B) > 0, the above formula
24 2 Some Probabilistic Concepts and Results

can be reversed to provide an expression for P(Aj|B), j = 1, 2, . . . . In


fact,

) ( P(B) ) = ( P(B) ) = (P(B A) )P( A ) .


P A B P B A P( A ) P B A P( A )
(
P Aj B =
j j j j j

i i i

Thus
THEOREM 5 (Bayes Formula) If {Aj, j = 1, 2, . . .} is a partition of S and P(Aj) > 0, j = 1,
2, . . . , and if P(B) > 0, then

(
P Aj B = ) (
)( ).
P B Aj P Aj

P( B A ) P( A )
i i i

REMARK 4 It is important that one checks to be sure that the collection


{Aj, j 1} forms a partition of S, as only then are the above theorems true.
The following simple example serves as an illustration of Theorems 4
and 5.
EXAMPLE 2 A multiple choice test question lists five alternative answers, of which only one
is correct. If a student has done the homework, then he/she is certain to
identify the correct answer; otherwise he/she chooses an answer at random.
Let p denote the probability of the event A that the student does the home-
work and let B be the event that he/she answers the question correctly. Find
the expression of the conditional probability P(A|B) in terms of p.
By noting that A and Ac form a partition of the appropriate sample space,
an application of Theorems 4 and 5 gives

( )
P AB =
( )( )
P BA P A
=
1 p
=
5p
.
( )( ) (
P BA P A +P BA P A c
)( ) c 1
(
1 p + 1 p
5
4p+1
)
Furthermore, it is easily seen that P(A|B) = P(A) if and only if p = 0 or 1.
For example, for p = 0.7, 0.5, 0.3, we find, respectively, that P(A|B) is
approximately equal to: 0.92, 0.83 and 0.68.
Of course, there is no reason to restrict ourselves to one partition of S
only. We may consider, for example, two partitions {Ai, i = 1, 2, . . .} {Bj, j = 1,
2, . . . }. Then, clearly,

(
Ai = Ai B j ,
j
) i = 1, 2, ,

Bj = ( B A ), j i j = 1, 2, ,
i

and

{A B , i = 1, 2, , j = 1, 2, }
i j
2.4 Combinatorial
Exercises
Results 25

is a partition of S. In fact,

(A B ) (A
i j i )
B j = if (i, j ) (i, j )
and

( A B ) = ( A B ) = A = S.
i j i j i
i, j i j i

The expression P(Ai Bj) is called the joint probability of Ai and Bj. On the
other hand, from
Ai = Ai Bj
j
( ) and Bj = Ai Bj ,
i
( )
we get

( ) (
P Ai = P Ai B j = P Ai B j P B j ,
j
) j
( )( )
provided P(Bj) > 0, j = 1, 2, . . . , and

( ) (
P B j = P Ai B j = P B j Ai P Ai ,
i
) i
( )( )
provided P(Ai) > 0, i = 1, 2, . . . . The probabilities P(Ai), P(Bj) are called
marginal probabilities. We have analogous expressions for the case of more
than two partitions of S.

Exercises
2.2.1 If P(A|B) > P(A), then show that P(B|A) > P(B) (P(A)P(B) > 0).
2.2.2 Show that:
i) P(Ac|B) = 1 P(A|B);
ii) P(A B|C) = P(A|C) + P(B|C) P(A B|C).
Also show, by means of counterexamples, that the following equations need
not be true:
iii) P(A|Bc) = 1 P(A|B);
iv) P(C|A + B) = P(C|A) + P(C|B).
2.2.3 If A B = and P(A + B) > 0, express the probabilities P(A|A + B)
and P(B|A + B) in terms of P(A) and P(B).
2.2.4 Use induction to prove Theorem 3.
2.2.5 Suppose that a multiple choice test lists n alternative answers of which
only one is correct. Let p, A and B be defined as in Example 2 and find Pn(A|B)
26 2 Some Probabilistic Concepts and Results

in terms of n and p. Next show that if p is fixed but different from 0 and 1, then
Pn(A|B) increases as n increases. Does this result seem reasonable?
2.2.6 If Aj, j = 1, 2, 3 are any events in S, show that {A1, Ac1 A2, Ac1 Ac2
A3, (A1 A2 A3)c} is a partition of S.
2.2.7 Let {Aj, j = 1, . . . , 5} be a partition of S and suppose that P(Aj) = j/15
and P(A|Aj) = (5 j)/15, j = 1, . . . , 5. Compute the probabilities P(Aj|A),
j = 1, . . . , 5.
2.2.8 A girls club has on its membership rolls the names of 50 girls with the
following descriptions:
20 blondes, 15 with blue eyes and 5 with brown eyes;
25 brunettes, 5 with blue eyes and 20 with brown eyes;
5 redheads, 1 with blue eyes and 4 with green eyes.
If one arranges a blind date with a club member, what is the probability that:
i) The girl is blonde?
ii) The girl is blonde, if it was only revealed that she has blue eyes?
2.2.9 Suppose that the probability that both of a pair of twins are boys is 0.30
and that the probability that they are both girls is 0.26. Given that the probabil-
ity of a child being a boy is 0.52, what is the probability that:
i) The second twin is a boy, given that the first is a boy?
ii) The second twin is a girl, given that the first is a girl?
2.2.10 Three machines I, II and III manufacture 30%, 30% and 40%, respec-
tively, of the total output of certain items. Of them, 4%, 3% and 2%, respec-
tively, are defective. One item is drawn at random, tested and found to be
defective. What is the probability that the item was manufactured by each one
of the machines I, II and III?
2.2.11 A shipment of 20 TV tubes contains 16 good tubes and 4 defective
tubes. Three tubes are chosen at random and tested successively. What is the
probability that:
i) The third tube is good, if the first two were found to be good?
ii) The third tube is defective, if one of the other two was found to be good
and the other one was found to be defective?
2.2.12 Suppose that a test for diagnosing a certain heart disease is 95%
accurate when applied to both those who have the disease and those who do
not. If it is known that 5 of 1,000 in a certain population have the disease in
question, compute the probability that a patient actually has the disease if the
test indicates that he does. (Interpret the answer by intuitive reasoning.)
2.2.13 Consider two urns Uj, j = 1, 2, such that urn Uj contains mj white balls
and nj black balls. A ball is drawn at random from each one of the two urns and
2.4 Combinatorial
2.3 Independence
Results 27

is placed into a third urn. Then a ball is drawn at random from the third urn.
Compute the probability that the ball is black.
2.2.14 Consider the urns of Exercise 2.2.13. A balanced die is rolled and if
an even number appears, a ball, chosen at random from U1, is transferred to
urn U2. If an odd number appears, a ball, chosen at random from urn U2, is
transferred to urn U1. What is the probability that, after the above experiment
is performed twice, the number of white balls in the urn U2 remains the same?
2.2.15 Consider three urns Uj, j = 1, 2, 3 such that urn Uj contains mj white
balls and nj black balls. A ball, chosen at random, is transferred from urn U1 to
urn U2 (color unnoticed), and then a ball, chosen at random, is transferred
from urn U2 to urn U3 (color unnoticed). Finally, a ball is drawn at random
from urn U3. What is the probability that the ball is white?
2.2.16 Consider the urns of Exercise 2.2.15. One urn is chosen at random
and one ball is drawn from it also at random. If the ball drawn was white, what
is the probability that the urn chosen was urn U1 or U2?
2.2.17 Consider six urns Uj, j = 1, . . . , 6 such that urn Uj contains mj ( 2)
white balls and nj ( 2) black balls. A balanced die is tossed once and if the
number j appears on the die, two balls are selected at random from urn Uj.
Compute the probability that one ball is white and one ball is black.
2.2.18 Consider k urns Uj, j = 1, . . . , k each of which contain m white balls
and n black balls. A ball is drawn at random from urn U1 and is placed in urn
U2. Then a ball is drawn at random from urn U2 and is placed in urn U3 etc.
Finally, a ball is chosen at random from urn Uk1 and is placed in urn Uk. A ball
is then drawn at random from urn Uk. Compute the probability that this last
ball is black.

2.3 Independence
For any events A, B with P(A) > 0, we defined P(B|A) = P(A B)/P(A). Now
P(B|A) may be >P(B), < P(B), or = P(B). As an illustration, consider an urn
containing 10 balls, seven of which are red, the remaining three being black.
Except for color, the balls are identical. Suppose that two balls are drawn
successively and without replacement. Then (assuming throughout the uni-
form probability function) the conditional probability that the second ball is
red, given that the first ball was red, is 69 , whereas the conditional probability
that the second ball is red, given that the first was black, is 79 . Without any
knowledge regarding the first ball, the probability that the second ball is red is
7
10
. On the other hand, if the balls are drawn with replacement, the probability
that the second ball is red, given that the first ball was red, is 107 . This probabil-
ity is the same even if the first ball was black. In other words, knowledge of the
event which occurred in the first drawing provides no additional information in
28 2 Some Probabilistic Concepts and Results

calculating the probability of the event that the second ball is red. Events like
these are said to be independent.
As another example, revisit the two-children families example considered
earlier, and define the events A and B as follows: A = children of both
genders, B = older child is a boy. Then P(A) = P(B) = P(B|A) = 12 . Again
knowledge of the event A provides no additional information in calculating
the probability of the event B. Thus A and B are independent.
More generally, let A, B be events with P(A) > 0. Then if P(B|A) = P(B),
we say that the even B is (statistically or stochastically or in the probability
sense) independent of the event A. If P(B) is also > 0, then it is easily seen that
A is also independent of B. In fact,

( ) (P(B) ) = ( P(B) ) = ( P)(B() ) = P( A).


P A B P B A P( A) P B P A
P AB =

That is, if P(A), P(B) > 0, and one of the events is independent of the other,
then this second event is also independent of the first. Thus, independence is
a symmetric relation, and we may simply say that A and B are independent. In
this case P(A B) = P(A)P(B) and we may take this relationship as the
definition of independence of A and B. That is,
DEFINITION 3 The events A, B are said to be (statistically or stochastically or in the probabil-
ity sense) independent if P(A B) = P(A)P(B).
Notice that this relationship is true even if one or both of P(A), P(B) = 0.
As was pointed out in connection with the examples discussed above,
independence of two events simply means that knowledge of the occurrence of
one of them helps in no way in re-evaluating the probability that the other
event happens. This is true for any two independent events A and B, as follows
from the equation P(A|B) = P(A), provided P(B) > 0, or P(B|A) = P(B),
provided P(A) > 0. Events which are intuitively independent arise, for exam-
ple, in connection with the descriptive experiments of successively drawing
balls with replacement from the same urn with always the same content, or
drawing cards with replacement from the same deck of playing cards, or
repeatedly tossing the same or different coins, etc.
What actually happens in practice is to consider events which are inde-
pendent in the intuitive sense, and then define the probability function P
appropriately to reflect this independence.
The definition of independence generalizes to any finite number of events.
Thus:
DEFINITION 4 The events Aj, j = 1, 2, . . . , n are said to be (mutually or completely) indepen-
dent if the following relationships hold:

( 1 k
) ( )
P Aj Aj = P Aj P Aj
1
( ) k

for any k = 2, . . . , n and j1, . . . , jk = 1, 2, . . . , n such that 1 j1 < j2 < < jk


n. These events are said to be pairwise independent if P(Ai Aj) = P(Ai)P(Aj)
for all i j.
2.4 Combinatorial
2.3 Independence
Results 29

It follows that, if the events Aj, j = 1, 2, . . . , n are mutually independent,


then they are pairwise independent. The converse need not be true, as the
example below illustrates. Also there are

n n n n n
+ + + =2 =2 n1
n n

2 3 n 1 0

relationships characterizing the independence of Aj, j = 1, . . . , n and they are


all necessary. For example, for n = 3 we will have:

( ) ( ) ( ) ( )
P A1 A2 A3 = P A1 P A2 P A3 ,
P( A A ) = P( A )P( A ),
1 2 1 2

P( A A ) = P( A )P( A ),
1 3 1 3

P( A A ) = P( A )P( A ).
2 3 2 3

That these four relations are necessary for the characterization of indepen-
dence of A1, A2, A3 is illustrated by the following examples:
Let S = {1, 2, 3, 4}, P({1}) = = P({4}) = 14 , and set A1 = {1, 2}, A2 = {1, 3},
A3 = {1, 4}. Then

A1 A2 = A1 A3 = A2 A3 = 1 , and {} A1 A2 A3 = 1 .{}
Thus

( ) ( ) (
P A1 A2 = P A1 A3 = P A2 A3 = P A1 A2 A3 = ) ( ) 1
4
.

Next,

(
P A1 A2 =
1 1 1
)
= = P A1 P A2 ,
4 2 2
( ) ( )
( 1 1 1
)
P A1 A3 = = = P A1 P A3 ,
4 2 2
( ) ( )
( 1 1 1
)
P A1 A3 = = = P A2 P A3 ,
4 2 2
( ) ( )
but

(
P A1 A2 A3 = ) 1 1 1 1
( ) ( ) ( )
= P A1 P A2 P A3 .
4 2 2 2

Now let S = {1, 2, 3, 4, 5}, and define P as follows:

({ })
P 1 =
1
8
, P 2 ({ }) = P({3}) = P({4}) = 163 , P({5}) = 165 .
30 2 Some Probabilistic Concepts and Results

Let
{ } {
A1 = 1, 2, 3 , A2 = 1, 2, 4 , A3 = 1, 3, 4 . } { }
Then
A1 A2 = 1, 2 , { } A1 A2 A3 = 1 . {}
Thus

(
P A1 A2 A3 = ) 1 1 1 1
= = P A1 P A2 P A3 ,
8 2 2 2
( ) ( ) ( )
but

(
P A1 A2 = ) 5 1 1
= P A1 P A2 .
16 2 2
( ) ( )
The following result, regarding independence of events, is often used by
many authors without any reference to it. It is the theorem below.
THEOREM 6 If the events A1, . . . , An are independent, so are the events A1, . . . , An, where
Aj is either Aj or Acj, j = 1, . . . , n.
PROOF The proof is done by (a double) induction. For n = 2, we have to
show that P(A1 A2) = P(A1)P(A2). Indeed, let A1 = A1 and A2 = Ac2. Then
P(A1 A2) = P(A1 Ac2) = P[A1 (S A2)] = P(A1 A1 A2) = P(A1)
P(A1 A2) = P(A1) P(A1)P(A2) = P(A1)[1 P(A2)] = P(A1)P(Ac2) = P(A1)P(A2).
Similarly if A1 = Ac1 and A2 = A2. For A1 = Ac1 and A2 = Ac2, P(A1 A2) =
P(Ac1 Ac2) = P[(S A1) Ac2] = P(Ac2 A1 Ac2) = P(Ac2) P(A1 Ac2) = P(Ac2)
P(A1)P(Ac2) = P(Ac2)[1 P(A1)] = P(Ac2)P(Ac1) = P(A1)P(A2).
Next, assume the assertion to be true for k events and show it to be true
for k + 1 events. That is, we suppose that P(A1 Ak) = P(A1 ) P(Ak),
and we shall show that P(A1 Ak+1) = P(A1 ) P(Ak+1). First, assume
that Ak+1 = Ak+1, and we have to show that

( ) (
P A1 Ak +1 = P A1 Ak Ak+1 )
= P( A) P( A )P( A ).
1 k k+1

This relationship is established also by induction as follows: If A1 = Ac1 and


Ai = Ai, i = 2, . . . , k, then

( ) [(
P A1c A2 Ak Ak+1 = P S A1 A2 Ak Ak+1 ) ]
(
= P A2 Ak Ak+1 A1 A2 Ak Ak+1 )
= P( A A A ) P( A A A A )
2 k k+1 1 2 k k+1

= P( A ) P( A ) P( A ) P( A ) P( A ) P( A ) P( A )
2 k k+1 1 2 k k+1

= P( A ) P( A )P( A )[1 P( A )] = P( A )P( A ) P( A )P( A ).


2 k k+1 1 1
c
2 k k+1
2.4 Combinatorial
2.3 Independence
Results 31

This is, clearly, true if Ac1 is replaced by any other Aci, i = 2, . . . , k. Now, for
! < k, assume that

(
P A1c Alc Al+1 Ak +1 )
( ) ( )( )
= P A1c P Alc P Al+1 P Ak +1 ( )
and show that

(
P A1c Alc Alc+1 Al+ 2 Ak +1 )
=P A ( ) P( A )P( A )P( A ) P( A ).
c
1
c
l
c
l +1 l+ 2 k +1

Indeed,

(
P A1c Alc Alc+1 Al+ 2 Ak+1 )
[ (
= P A1c Alc S Al+1 Al+ 2 Ak+1 ) ]
(
= P A A Al+ 2 Ak+1
1
c c
l

A1c Alc Al+1 Al+ 2 Ak+1 )


(
= P A A Al+ 2 Ak+1
1
c c
l )
(
P A A Al+1 Al+ 2 Ak+1
1
c c
l )
( ) ( )( ) ( )
= P A1c P Alc P Al+ 2 P Ak+1

P( A ) P( A ) P( A ) P( A ) P( A )
1
c c
l l +1 l+ 2 k+1

(by the induction hypothesis)


= P( A ) P( A )P( A ) P( A )[1 P( A )]
1
c c
l l+ 2 k+1 l +1

= P( A ) P( A ) P( A ) P( A ) P( A )
1
c c
l l+ 2 k+1
c
l +1

= P( A ) P( A )P( A )P( A ) P( A ),
1
c c
l
c
l +1 l+ 2 k+1

as was to be seen. It is also, clearly, true that the same result holds if the ! Ai s
which are Aci are chosen in any one of the (k!) possible ways of choosing ! out
of k. Thus, we have shown that

(
P A1 Ak Ak+1 = P A1 P Ak P Ak+1 . ) ( ) ( ) ( )
Finally, under the assumption that

( )
P A1 Ak = P A1 P Ak , ( ) ( )
take Ak+1 = A , and show that
c
k+1

(
P A1 Ak Akc+1 = P A1 P Ak P Akc+1 .) ( ) ( ) ( )
32 2 Some Probabilistic Concepts and Results

In fact,

( ) [(
P A1 Ak Akc+1 = P A1 Ak S Ak+1 ( ))]
(
= P A1 Ak A1 Ak Ak+1 )
= P( A A ) P( A A A )
1 k 1 k k+1

= P( A) P( A ) P( A) P( A )P( A )
1 k 1 k k+1

(by the induction hypothesis and what was last proved)


= P( A) P( A )[1 P( A )]
1 k k+1

= P( A) P( A )P( A ).
1 k
c
k+1

This completes the proof of the theorem.


Now, for j = 1, 2, let Ej be an experiment having the sample space Sj. One
may look at the pair (E1, E2) of experiments, and then the question arises as to
what is the appropriate sample space for this composite or compound experi-
ment, also denoted by E1 E2. If S stands for this sample space, then, clearly,
S = S1 S2 = {(s1, s2); s1 S1, s2 S2}. The corresponding events are, of course,
subsets of S. The notion of independence also carries over to experiments.
Thus, we say that the experiments E1 and E2 are independent if P(B1 B2) =
P(B1)P(B2) for all events B1 associated with E1 alone, and all events B2 associ-
ated E2 alone.
What actually happens in practice is to start out with two experiments E1,
E2 which are intuitively independent, such as the descriptive experiments (also
mentioned above) of successively drawing balls with replacement from the
same urn with always the same content, or drawing cards with replacement
from the same deck of playing cards, or repeatedly tossing the same or differ-
ent coins etc., and have the corresponding probability spaces (S1, class of
events, P1) and (S2, class of events, P2), and then define the probability func-
tion P, in terms of P1 and P2, on the class of events in the space S1 S2 so that
it reflects the intuitive independence.
The above definitions generalize in a straightforward manner to any finite
number of experiments. Thus, if Ej, j = 1, 2, . . . , n, are n experiments with
corresponding sample spaces Sj and probability functions Pj on the respective
classes of events, then the compound experiment
(E , E , , E ) = E
1 2 n 1 E2 En
has sample space S, where
S = S1 S n = {(s , , s ); s S , j = 1, 2, , n}.
1 n j j

The class of events are subsets of S, and the experiments are said to be
independent if for all events Bj associated with experiment Ej alone, j = 1,
2, . . . , n, it holds
2.4 Combinatorial
Exercises
Results 33

( ) ( ) ( )
P B1 Bn = P B1 P Bn .
Again, the probability function P is defined, in terms of Pj, j = 1, 2, . . . , n, on
the class of events in S so that to reflect the intuitive independence of the
experiments Ej, j = 1, 2, . . . , n.
In closing this section, we mention that events and experiments which are
not independent are said to be dependent.

Exercises
2.3.1 If A and B are disjoint events, then show that A and B are independent
if and only if at least one of P(A), P(B) is zero.
2.3.2 Show that if the event A is independent of itself, then P(A) = 0 or 1.
2.3.3 If A, B are independent, A, C are independent and B C = , then A,
B + C are independent. Show, by means of a counterexample, that the conclu-
sion need not be true if B C .
2.3.4 For each j = 1, . . . , n, suppose that the events A1, . . . , Am, Bj are
independent and that Bi Bj = , i j. Then show that the events A1, . . . , Am,
nj= 1Bj are independent.
2.3.5 If Aj, j = 1, . . . , n are independent events, show that
n
( )
n
P U A j = 1 P Acj .
j =1 j =1

2.3.6 Jim takes the written and road drivers license tests repeatedly until he
passes them. Given that the probability that he passes the written test is 0.9
and the road test is 0.6 and that tests are independent of each other, what is the
probability that he will pass both tests on his nth attempt? (Assume that
the road test cannot be taken unless he passes the written test, and that once
he passes the written test he does not have to take it again, no matter whether
he passes or fails his next road test. Also, the written and the road tests are
considered distinct attempts.)
2.3.7 The probability that a missile fired against a target is not intercepted by
an antimissile missile is 32 . Given that the missile has not been intercepted, the
probability of a successful hit is 34 . If four missiles are fired independently,
what is the probability that:
i) All will successfully hit the target?
ii) At least one will do so?
How many missiles should be fired, so that:
34 2 Some Probabilistic Concepts and Results

iii) At least one is not intercepted with probability 0.95?


iv) At least one successfully hits its target with probability 0.99?
2.3.8 Two fair dice are rolled repeatedly and independently. The first time a
total of 10 appears, player A wins, while the first time that a total of 6 appears,
player B wins, and the game is terminated. Compute the probabilities that:
i) The game terminates on the nth throw and player A wins;
ii) The same for player B;
iii) Player A wins;
iv) Player B wins;
v) Does the game terminate ever?
2.3.9 Electric current is transmitted from point A to point B provided at
least one of the circuits #1 through #n below is closed. If the circuits close
independently of each other and with respective probabilities pi, i = 1, . . . , n,
determine the probability that:
i) Exactly one circuit is closed;
ii) At least one circuit is closed;
iii) Exactly m circuits are closed for 0 m n;
iv) At least m circuits are closed with m as in part (iii);
v) What do parts (i)(iv) become for p1 = = pn = p, say?

A B

2.4 Combinatorial Results


In this section, we will restrict ourselves to finite sample spaces and uniform
probability functions. Some combinatorial results will be needed and we pro-
ceed to derive them here. Also examples illustrating the theorems of previous
sections will be presented.
2.4 Combinatorial Results 35

The following theorem, known as the Fundamental Principle of Counting,


forms the backbone of the results in this section.
THEOREM 7 Let a task T be completed by carrying out all of the subtasks Tj, j = 1, 2, . . . ,
k, and let it be possible to perform the subtask Tj in nj (different) ways, j = 1,
2, . . . , k. Then the total number of ways the task T may be performed is given
by kj = 1 n j .
PROOF The assertion is true for k = 2, since by combining each one of the n1
ways of performing subtask T1 with each one of the n2 ways of performing
subtask T2, we obtain n1n2 as the total number of ways of performing task T.
Next, assume the result to be true for k = m and establish it for k = m + 1. The
reasoning is the same as in the step just completed, since by combining each
one of the mj = 1 n j ways of performing the first m subtasks with each one of
( )
nm+1 ways of performing substask Tm+1, we obtain mj = 1 n j nm+1 = jm=+11 n j for
the total number of ways of completing task T.
The following examples serve as an illustration to Theorem 7.
EXAMPLE 3 i) A man has five suits, three pairs of shoes and two hats. Then the number
of different ways he can attire himself is 5 3 2 = 30.
ii) Consider the set S = {1, . . . , N} and suppose that we are interested in
finding the number of its subsets. In forming a subset, we consider for each
element whether to include it or not. Then the required number is equal to
the following product of N factors 2 2 = 2N.
iii) Let nj = n(Sj) be the number of points of the sample space Sj, j = 1, 2, . . . ,
k. Then the sample space S = S1 Sk has n(S) = n1 nk sample
points. Or, if nj is the number of outcomes of the experiment Ej, j = 1,
2, . . . , k, then the number of outcomes of the compound experiment E1
. . . Ek is n1 . . . nk.
In the following, we shall consider the problems of selecting balls from an
urn and also placing balls into cells which serve as general models of
many interesting real life problems. The main results will be formulated as
theorems and their proofs will be applications of the Fundamental Principle of
Counting.
Consider an urn which contains n numbered (distinct, but otherwise iden-
tical) balls. If k balls are drawn from the urn, we say that a sample of size k was
drawn. The sample is ordered if the order in which the balls are drawn is taken
into consideration and unordered otherwise. Then we have the following
result.
THEOREM 8 i) The number of ordered samples of size k is n(n 1) (n k + 1) = Pn,k
(permutations of k objects out of n, and in particular, if k = n, Pn,n = 1 2
n = n!), provided the sampling is done without replacement; and is equal to
nk if the sampling is done with replacement.
ii) The number of unordered samples of size k is
36 2 Some Probabilistic Concepts and Results

Pn ,k n n!
= C n ,k = =
k! (
k k! n k !)
if the sampling is done without replacement; and is equal to
n + k 1
( )
N n, k =
k

if the sampling is done with replacement. [See also Theorem 9(iii).]


PROOF
i) The first part follows from Theorem 7 by taking nj = (n j + 1), j = 1, . . . ,
k, and the second part follows from the same theorem by taking nj = n,
j = 1, . . . , k.
ii) For the first part, we have that, if order counts, this number is Pn,k. Since for
every sample of size k one can form k! ordered samples of the same size, if
x is the required number, then Pn,k = xk!. Hence the desired result.
The proof of the second part may be carried out by an appropriate induc-
tion method. However, we choose to present the following short alternative
proof which is due to S. W. Golomb and appeared in the American Mathemati-
cal Monthly, 75, 1968, p. 530. For clarity, consider the n balls to be cards
numbered from 1 to n and adjoin k 1 extra cards numbered from n + 1 to
n + k 1 and bearing the respective instructions: repeat lowest numbered
card, repeat 2nd lowest numbered card, . . . , repeat (k 1)st lowest num-
bered card. Then a sample of size k without replacement from this enlarged
(n + k 1)-card deck corresponds uniquely to a sample of size k from the
original deck with replacement. (That is, take k out of n + k 1, without
replacement so that there will be at least one out of 1, 2, . . . , n, and then apply
the instructions.) Thus, by the first part, the required number is

n + k 1

k
(
= N n, k , )
as was to be seen.
For the sake of illustration of Theorem 8, let us consider the following
examples.

EXAMPLE 4 i(i) Form all possible three digit numbers by using the numbers 1, 2, 3, 4, 5.
(ii) Find the number of all subsets of the set S = {1, . . . , N}.
In part (i), clearly, the order in which the numbers are selected is relevant.
Then the required number is P5,3 = 5 4 3 = 60 without repetitions, and 53 = 125
with repetitions.
In part (ii) the order is, clearly, irrelevant and the required number is (N0)
+ ( 1) + + (NN) = 2N, as already found in Example 3.
N
2.4 Combinatorial Results 37

EXAMPLE 5 An urn contains 8 balls numbered 1 to 8. Four balls are drawn. What is the
probability that the smallest number is 3?
Assuming the uniform probability function, the required probabilities are
as follows for the respective four possible sampling cases:

5

3 1
Order does not count/replacements not allowed: = 0.14;
8 7

4

6 + 3 1

3 28
Order does not count/replacements allowed: = 0.17;
8 + 4 1 165

4

Order counts/replacements not allowed:


(5 4 3)4 = 1 0.14;
8765 7
4 3 4 2 4 4
1 5 + 2 5 + 3 5 + 4
Order counts/replacements allowed:
=
671
0.16.
84 4, 096
EXAMPLE 6 What is the probability that a poker hand will have exactly one pair?
A poker hand is a 5-subset of the set of 52 cards in a full deck, so there
are
52
= N = 2, 598, 960
5
different poker hands. We thus let S be a set with N elements and assign the
uniform probability measure to S. A poker hand with one pair has two cards
of the same face value and three cards whose faces are all different among
themselves and from that of the pair. We arrive at a unique poker hand with
one pair by completing the following tasks in order:
a) Choose the face value of the pair from the 13 available face values. This can
be done in (131) = 13 ways.
b) Choose two cards with the face value selected in (a). This can be done in (42)
= 6 ways.
c) Choose the three face values for the other three cards in the hand. Since
there are 12 face values to choose from, this can be done in (123) = 220
ways.
d) Choose one card (from the four at hand) of each face value chosen in (c).
This can be done in 4 4 4 = 43 = 64 ways.
38 2 Some Probabilistic Concepts and Results

Then, by Theorem 6, there are 13 6 220 64 = 1,098,240 poker hands with


one pair. Hence, by assuming the uniform probability measure, the required
probability is equal to

1, 098, 240
0.42.
2, 598, 960

THEOREM 9 i) The number of ways in which n distinct balls can be distributed into k
distinct cells is kn.
ii) The number of ways that n distinct balls can be distributed into k distinct
cells so that the jth cell contains nj balls (nj 0, j = 1, . . . , k, kj=1 nj = n)
is

n! n
= .
n1! n2 ! nk ! n1 , n2 , , nk

iii) The number of ways that n indistinguishable balls can be distributed into
k distinct cells is

k + n 1
.
n

Furthermore, if n k and no cell is to be empty, this number becomes

n 1
.
k 1

PROOF
i) Obvious, since there are k places to put each of the n balls.
ii) This problem is equivalent to partitioning the n balls into k groups, where
the jth group contains exactly nj balls with nj as above. This can be done in
the following number of ways:

n n n1 n n1 nk1 n!
= n !n ! n !.
n1 n2 nk 1 2 k

iii) We represent the k cells by the k spaces between k + 1 vertical bars and the
n balls by n stars. By fixing the two extreme bars, we are left with k + n
1 bars and stars which we may consider as k + n 1 spaces to be filled in
by a bar or a star. Then the problem is that of selecting n spaces for the n
( )
stars which can be done in k+nn1 ways. As for the second part, we now
have the condition that there should not be two adjacent bars. The n stars

k 1 bars, the result follows.


k1 ( )
create n 1 spaces and by selecting k 1 of them in n1 ways to place the
2.4 Combinatorial Results 39

REMARK 5
i) The numbers nj, j = 1, . . . , k in the second part of the theorem are called
occupancy numbers.
ii) The answer to (ii) is also the answer to the following different question:
Consider n numbered balls such that nj are identical among themselves and
distinct from all others, nj 0, j = 1, . . . , k, kj=1 nj = n. Then the number of
different permutations is
n
.
n1 , n2 , , nk
Now consider the following examples for the purpose of illustrating the
theorem.

EXAMPLE 7 Find the probability that, in dealing a bridge hand, each player receives one
ace.
The number of possible bridge hands is
52 52!
N = = .
13, 13, 13, 13 ( )
4
13!

Our sample space S is a set with N elements and assign the uniform probability
measure. Next, the number of sample points for which each player, North,
South, East and West, has one ace can be found as follows:
a) Deal the four aces, one to each player. This can be done in
4 4!
= 1! 1! 1! 1! = 4! ways.
1, 1, 1, 1
b) Deal the remaining 48 cards, 12 to each player. This can be done in

48 48!
= ways.
12, 12, 12, 12 ( )
4
12!

Thus the required number is 4!48!/(12!)4 and the desired probability is


4!48!(13!)4/[(12!)4(52!)]. Furthermore, it can be seen that this probability lies
between 0.10 and 0.11.

EXAMPLE 8 The eleven letters of the word MISSISSIPPI are scrambled and then arranged
in some order.
i) What is the probability that the four Is are consecutive letters in the
resulting arrangement?
There are eight possible positions for the first I and the remaining

probability is
( )
seven letters can be arranged in 7 distinct ways. Thus the required
1, 4 , 2
40 2 Some Probabilistic Concepts and Results

7
8
1, 4, 2 4
= 0.02.
11 165

1, 4, 4, 2
ii) What is the conditional probability that the four Is are consecutive (event
A), given B, where B is the event that the arrangement starts with M and
ends with S?
Since there are only six positions for the first I, we clearly have

5
6
( )
P AB =
2
9
=
1
21
0.05.

4, 3, 2
iii) What is the conditional probability of A, as defined above, given C, where
C is the event that the arrangement ends with four consecutive Ss?
Since there are only four positions for the first I, it is clear that

3
4
( )
P AC =
2
7
=
4
35
0.11.

1, 2, 4

Exercises
2.4.1 A combination lock can be unlocked by switching it to the left and
stopping at digit a, then switching it to the right and stopping at digit b and,
finally, switching it to the left and stopping at digit c. If the distinct digits a, b
and c are chosen from among the numbers 0, 1, . . . , 9, what is the number of
possible combinations?
2.4.2 How many distinct groups of n symbols in a row can be formed, if each
symbol is either a dot or a dash?
2.4.3 How many different three-digit numbers can be formed by using the
numbers 0, 1, . . . , 9?
2.4.4 Telephone numbers consist of seven digits, three of which are grouped
together, and the remaining four are also grouped together. How many num-
bers can be formed if:
i) No restrictions are imposed?
ii) If the first three numbers are required to be 752?
2.4 Combinatorial Results
Exercises 41

2.4.5 A certain state uses five symbols for automobile license plates such that
the first two are letters and the last three numbers. How many license plates
can be made, if:
i) All letters and numbers may be used?
ii) No two letters may be the same?
2.4.6 Suppose that the letters C, E, F, F, I and O are written on six chips and
placed into an urn. Then the six chips are mixed and drawn one by one without
replacement. What is the probability that the word OFFICE is formed?
2.4.7 The 24 volumes of the Encyclopaedia Britannica are arranged on a
shelf. What is the probability that:
i) All 24 volumes appear in ascending order?
ii) All 24 volumes appear in ascending order, given that volumes 14 and 15
appeared in ascending order and that volumes 113 precede volume 14?
2.4.8 If n countries exchange ambassadors, how many ambassadors are
involved?
2.4.9 From among n eligible draftees, m men are to be drafted so that all
possible combinations are equally likely to be chosen. What is the probability
that a specified man is not drafted?
2.4.10 Show that

n + 1

m + 1 n+1
= .
n m+1

m
2.4.11 Consider five line segments of length 1, 3, 5, 7 and 9 and choose three
of them at random. What is the probability that a triangle can be formed by
using these three chosen line segments?
2.4.12 From 10 positive and 6 negative numbers, 3 numbers are chosen at
random and without repetitions. What is the probability that their product is
a negative number?
2.4.13 In how many ways can a committee of 2n + 1 people be seated along
one side of a table, if the chairman must sit in the middle?
2.4.14 Each of the 2n members of a committee flips a fair coin in deciding
whether or not to attend a meeting of the committee; a committee member
attends the meeting if an H appears. What is the probability that a majority
will show up in the meeting?
2.4.15 If the probability that a coin falls H is p (0 < p < 1), what is the
probability that two people obtain the same number of Hs, if each one of
them tosses the coin independently n times?
42 2 Some Probabilistic Concepts and Results

2.4.16

i) Six fair dice are tossed once. What is the probability that all six faces
appear?
ii) Seven fair dice are tossed once. What is the probability that every face
appears at least once?
2.4.17 A shipment of 2,000 light bulbs contains 200 defective items and 1,800
good items. Five hundred bulbs are chosen at random, are tested and the
entire shipment is rejected if more than 25 bulbs from among those tested are
found to be defective. What is the probability that the shipment will be
accepted?
2.4.18 Show that
M M 1 M 1
= + ,
m m m 1
where N, m are positive integers and m < M.
2.4.19 Show that
r
m n m + n
x r x = r
,
x =0

where
k
= 0 if x > k.
x
2.4.20 Show that
n
n n
n
(1) j = 0.
j
i) j = 2 n ; ii )
j=0 j=0

2.4.21 A student is given a test consisting of 30 questions. For each question


there are supplied 5 different answers (of which only one is correct). The
student is required to answer correctly at least 25 questions in order to pass the
test. If he knows the right answers to the first 20 questions and chooses an
answer to the remaining questions at random and independently of each other,
what is the probability that he will pass the test?
2.4.22 A student committee of 12 people is to be formed from among 100
freshmen (60 male + 40 female), 80 sophomores (50 male + 30 female), 70
juniors (46 male + 24 female), and 40 seniors (28 male + 12 female). Find the
total number of different committees which can be formed under each one of
the following requirements:
i) No restrictions are imposed on the formation of the committee;
ii) Seven students are male and five female;
2.4 Combinatorial Results
Exercises 43

iii) The committee contains the same number of students from each class;
iv) The committee contains two male students and one female student from
each class;
v) The committee chairman is required to be a senior;
vi) The committee chairman is required to be both a senior and male;
vii) The chairman, the secretary and the treasurer of the committee are all
required to belong to different classes.
2.4.23 Refer to Exercise 2.4.22 and suppose that the committee is formed by
choosing its members at random. Compute the probability that the committee
to be chosen satisfies each one of the requirements (i)(vii).
2.4.24 A fair die is rolled independently until all faces appear at least once.
What is the probability that this happens on the 20th throw?

2.4.25 Twenty letters addressed to 20 different addresses are placed at ran-


dom into the 20 envelopes. What is the probability that:
i) All 20 letters go into the right envelopes?
ii) Exactly 19 letters go into the right envelopes?
iii) Exactly 17 letters go into the right envelopes?
2.4.26 Suppose that each one of the 365 days of a year is equally likely to be
the birthday of each one of a given group of 73 people. What is the probability
that:
i) Forty people have the same birthdays and the other 33 also have the same
birthday (which is different from that of the previous group)?
ii) If a year is divided into five 73-day specified intervals, what is the probabil-
ity that the birthday of: 17 people falls into the first such interval, 23 into
the second, 15 into the third, 10 into the fourth and 8 into the fifth interval?

2.4.27 Suppose that each one of n sticks is broken into one long and one
short part. Two parts are chosen at random. What is the probability that:
i) One part is long and one is short?
ii) Both parts are either long or short?
The 2n parts are arranged at random into n pairs from which new sticks are
formed. Find the probability that:
iii) The parts are joined in the original order;
iv) All long parts are paired with short parts.
2.4.28 Derive the third part of Theorem 9 from Theorem 8(ii).
2.4.29 Three cards are drawn at random and with replacement from a stan-
dard deck of 52 playing cards. Compute the probabilities P(Aj), j = 1, . . . , 5,
where the events Aj, j = 1, . . . , 5 are defined as follows:
44 2 Some Probabilistic Concepts and Results

{
A1 = s S ; all 3 cards in s are black , }
A2 = {s S ; at least 2 cards in s are red},
A3 = {s S ; exactly 1 card in s is an ace},
A4 = {s S ; the first card in s is a diamond,
}
the second is a heart and the third is a club ,

{ }
A5 = s S ; 1 card in s is a diamond, 1 is a heart and 1 is a club .

2.4.30 Refer to Exercise 2.4.29 and compute the probabilities P(Aj),


j = 1, . . . , 5 when the cards are drawn at random but without replacement.
2.4.31 Consider hands of 5 cards from a standard deck of 52 playing
cards. Find the number of all 5-card hands which satisfy one of the following
requirements:
i) Exactly three cards are of one color;
ii) Three cards are of three suits and the other two of the remaining suit;
iii) At least two of the cards are aces;
iv) Two cards are aces, one is a king, one is a queen and one is a jack;
v) All five cards are of the same suit.
2.4.32 An urn contains nR red balls, nB black balls and nW white balls. r balls
are chosen at random and with replacement. Find the probability that:
i) All r balls are red;
ii) At least one ball is red;
iii) r1 balls are red, r2 balls are black and r3 balls are white (r1 + r2 + r3 = r);
iv) There are balls of all three colors.
2.4.33 Refer to Exercise 2.4.32 and discuss the questions (i)(iii) for r = 3 and
r1 = r2 = r3 (= 1), if the balls are drawn at random but without replacement.
2.4.34 Suppose that all 13-card hands are equally likely when a standard
deck of 52 playing cards is dealt to 4 people. Compute the probabilities P(Aj),
j = 1, . . . , 8, where the events Aj, j = 1, . . . , 8 are defined as follows:

{
A1 = s S ; s consists of 1 color cards , }
A2 = {s S ; s consists only of diamonds},
A3 = {s S ; s consists of 5 diamonds, 3 hearts, 2 clubs and 3 spades},
A4 = {s S ; s consists of cards of exactly 2 suits},
A5 = {s S ; s contains at least 2 aces},
A6 = {s S ; s does not contain aces, tens and jacks},
2.5 Product
2.4 Combinatorial
Probability Results
Spaces 45

{
A7 = s S ; s consists of 3 aces, 2 kings and exactly 7 red cards , }
A8 = {s S ; s consists of cards of all different denominations}.

2.4.35 Refer to Exercise 2.4.34 and for j = 0, 1, . . . , 4, define the events Aj


and also A as follows:

{ }
A j = s S ; s contains exactly j tens ,
A = {s S ; s contains exactly 7 red cards}.

For j = 0, 1, . . . , 4, compute the probabilities P(Aj), P(Aj|A) and also P(A);


compare the numbers P(Aj), P(Aj|A).
2.4.36 Let S be the set of all n3 3-letter words of a language and let P be the
equally likely probability function on the events of S. Define the events A, B
and C as follows:

{
A = s S ; s begins with a specific letter , }
{ (
B = s S ; s has the specified letter mentioned in the definition of A )
in the middle entry , }
{
C = s S ; s has exactly two of its letters the same . }
Then show that:
i) P(A B) = P(A)P(B);
ii) P(A C) = P(A)P(C);
iii) P(B C) = P(B)P(C);
iv) P(A B C) P(A)P(B)P(C).
Thus the events A, B, C are pairwise independent but not mutually
independent.

2.5* Product Probability Spaces


The concepts discussed in Section 2.3 can be stated precisely by utilizing more
technical language. Thus, if we consider the experiments E1 and E2 with re-
spective probability spaces (S1, A1, P1) and (S2, A2, P2), then the compound
experiment (E1, E2) = E1 E2 has sample space S = S1 S2 as defined earlier.
The appropriate -field A of events in S is defined as follows: First define the
class C by:

{
C = A1 A2 ; A1 A1 , A2 A2 , }
where A1 A2 = {(s , s ); s A , s
1 2 1 1 2 }
A2 .
46 2 Some Probabilistic Concepts and Results

Then A is taken to be the -field generated by C (see Theorem 4 in Chapter


1). Next, define on C the set function P by P(A1 A2) = P1(A1)P2(A2). It can be
shown that P determines uniquely a probability measure on A (by means of
the so-called Carathodory extension theorem). This probability measure is
usually denoted by P1 P2 and is called the product probability measure (with
factors P1 and P2), and the probability space (S, A, P) is called the product
probability space (with factors (Sj, Aj, Pj), j = 1, 2). It is to be noted that events
which refer to E1 alone are of the form B1 = A1 S2, A1 A1, and those
referring to E2 alone are of the form B2 = S1 A2, A2 A2. The experiments E1
and E2 are then said to be independent if P(B1 B2) = P(B1)P(B2) for all events
B1 and B2 as defined above.
For n experiments Ej, j = 1, 2, . . . , n with corresponding probability spaces
(Sj, Aj, Pj), the compound experiment (E1, . . . , En) = E1 En has prob-
ability space (S, A, P), where
S = S1 S n = {(s , , s ); s S , j = 1, 2, , n},
1 n j j

A is the -field generated by the class C, where

{
C = A1 An ; A j A j , j = 1, 2, , n , }
and P is the unique probability measure defined on A through the
relationships

( ) ( ) ( )
P A1 An = P A1 P An , A j A j , j = 1, 2, , n.
The probability measure P is usually denoted by P1 Pn and is called the
product probability measure (with factors Pj, j = 1, 2, . . . , n), and the probabil-
ity space (S, A, P) is called the product probability space (with factors (Sj, Aj,
Pj), j = 1, 2, . . . , n). Then the experiments Ej, j = 1, 2, . . . , n are said to be
independent if P(B1 B2) = P(B1) P(B2), where Bj is defined by

B j = S1 S j 1 A j S j + 1 S n , j = 1, 2, , n.

The definition of independent events carries over to -fields as follows.


Let A1, A2 be two sub--fields of A. We say that A1, A2 are independent if
P(A1 A2) = P(A1)P(A2) for any A1 A1, A2 A2. More generally, the
-fields Aj, j = 1, 2, . . . , n (sub--fields of A) are said to be independent if
n n
P I A j = P A j
j =1 j =1
( ) for any A j A j , j = 1, 2, , n.

Of course, -fields which are not independent are said to be dependent.


At this point, notice that the factor -fields Aj, j = 1, 2, . . . , n may be
considered as sub--fields of the product -field A by identifying Aj with Bj,
where the Bjs are defined above. Then independence of the experiments Ej,
j = 1, 2, . . . , n amounts to independence of the corresponding -fields Aj,
j = 1, 2, . . . , n (looked upon as sub--fields of the product -field A).
2.6* The
2.4Probability
Combinatorial
of Matchings
Results 47

Exercises
2.5.1 Form the Cartesian products A B, A C, B C, A B C, where
A = {stop, go}, B = {good, defective), C = {(1, 1), (1, 2), (2, 2)}.

2.5.2 Show that A B = if and only if at least one of the sets A, B is .

2.5.3 If A B, show that A C B C for any set C.

2.5.4 Show that


i) (A B)c = (A Bc) + (Ac B) + (Ac Bc);
ii) (A B) (C D) = (A C) (B D);
iii) (A B) (C D) = (A C) (B D) [(A Cc) (Bc D)
+ (Ac C) (B Dc )].

2.6* The Probability of Matchings


In this section, an important result, Theorem 10, is established providing an
expression for the probability of occurrence of exactly m events out of possible
M events. The theorem is then illustrated by means of two interesting exam-
ples. For this purpose, some additional notation is needed which we proceed to
introduce. Consider M events Aj, j = 1, 2, . . . , M and set

S0 = 1,
M
S1 = P A j ,
j =1
( )
S2 =
1 j1 < j2 M
(
P A j1 A j2 , )
M
Sr =
1 j1 < j2 < < jr M
( )
P A j1 A j2 A jr ,

M
(
S M = P A1 A2 AM . )
Let also

Bm = exactly

C m = at least m of the events A j , j = 1, 2, , M occur.
Dm = at most

Then we have
48 2 Some Probabilistic Concepts and Results

THEOREM 10 With the notation introduced above


m + 1 m + 2 M
( ) ( )
M m
P Bm = S m S m+1 + S m+ 2 + 1 SM (2)
m m m
which for m = 0 is

( ) ( )
M
P B0 = S0 S1 + S 2 + 1 SM , (3)
and

( ) ( ) ( )
P C m = P Bm + P Bm+1 + + P BM , ( ) (4)
and

( ) ( ) ( )
P Dm = P B0 + P B1 + + P Bm . ( ) (5)

For the proof of this theorem, all that one has to establish is (2), since (4)
and (5) follow from it. This will be done in Section 5.6 of Chapter 5. For a proof
where S is discrete the reader is referred to the book An Introduction to
Probability Theory and Its Applications, Vol. I, 3rd ed., 1968, by W. Feller, pp.
99100.
The following examples illustrate the above theorem.

EXAMPLE 9 The matching problem (case of sampling without replacement). Suppose that
we have M urns, numbered 1 to M. Let M balls numbered 1 to M be inserted
randomly in the urns, with one ball in each urn. If a ball is placed into the urn
bearing the same number as the ball, a match is said to have occurred.
i) Show the probability of at least one match is

1 1
( ) 1
M +1
1 + + 1 1 e 1 0.63
2! 3! M!
for large M, and
ii) exactly m matches will occur, for m = 0, 1, 2, . . . , M is

1
1 1
( ) 1
M m
1 1 + + + 1
m! 2! 3! ( )
M m !
1 M m
( ) 1 1 1
k
= 1
m! k = 0

k! m!
e for M m large.

DISCUSSION To describe the distribution of the balls among the urns, write
an M-tuple (z1, z2, . . . , zM) whose jth component represents the number of the
ball inserted in the jth urn. For k = 1, 2, . . . , M, the event Ak that a match will
occur in the kth urn may be written Ak = {(z1, . . . , zM) ! M; zj integer, 1 zj
2.6* The
2.4Probability of Matchings
Combinatorial Results 49

M, j = 1, . . . , M, zk = k}. It is clear that for any integer r = 1, 2, . . . , M and any


r unequal integers k1, k2, . . . , kr,, from 1 to M,

) ( M! ) .
M r !
(
P Ak1 Ak2 Akr =

It then follows that Sr is given by

( )
M M r ! 1
Sr = = .
r M! r!
This implies the desired results.
EXAMPLE 10 Coupon collecting (case of sampling with replacement). Suppose that a manu-
facturer gives away in packages of his product certain items (which we take to
be coupons), each bearing one of the integers 1 to M, in such a way that each
of the M items is equally likely to be found in any package purchased. If n
packages are bought, show that the probability that exactly m of the integers,
1 to M, will not be obtained is equal to
n
M M m k M m m + k
( )
1
m k = 0 k
1 M .

Many variations and applications of the above problem are described in
the literature, one of which is the following. If n distinguishable balls are
distributed among M urns, numbered 1 to M, what is the probability that there
will be exactly m urns in which no ball was placed (that is, exactly m urns
remain empty after the n balls have been distributed)?
DISCUSSION To describe the coupons found in the n packages purchased,
we write an n-tuple (z1, z2, , zn), whose jth component zj represents the
number of the coupon found in the jth package purchased. We now define the
events A1, A2, . . . , AM. For k = 1, 2, . . . , M, Ak is the event that the number k
will not appear in the sample, that is,


( )
Ak = z1 , , zn ! n ; z j integer, 1 z j M , z j k, j = 1, 2, , n.

It is easy to see that we have the following results:
n n
M 1 1
( )
P Ak =
M
= 1 ,
M
k = 1, 2, , M ,

n n
M 2 2 k1 = 1, 2, , n
(
P Ak Ak1 2
) =
M
= 1 ,
M k2 = k1 + 1, , n

and, in general,
50 2 Some Probabilistic Concepts and Results

k1 = 1, 2, , n
n
r k2 = k1 + 1, , n
( 1 2 r
M
)
P Ak Ak Ak = 1 ,
M
kr = kr 1 + 1, , n.

Thus the quantities Sr are given by


n
M r
S r = 1 , r = 0, 1, , M . (6)
r M

Let Bm be the event that exactly m of the integers 1 to M will not be found in
the sample. Clearly, Bm is the event that exactly m of the events A1, . . . , AM
will occur. By relations (2) and (6), we have
n
M
r M r
( ) ( )
r m
P Bm = 1 1 M
r =m m r
n
M M m k M m m + k
= 1
m k=0 k
( )
1 M ,

(7)

by setting r m = k and using the identity

m + k M M M m
= . (8)
m m + k m k

This is the desired result.


This section is concluded with the following important result stated as a
theorem.
THEOREM 11 Let A and B be two disjoint events. Then in a series of independent trials, show
that:

( )
P A
(
P A occurs before B occurs = ) .
( ) ( )
P A +P B

PROOF For i = 1, 2, . . . , define the events Ai and Bi as follows:

Ai = A occurs on the ith trial, Bi = B occurs on the ith trial.


Then, clearly, required the event is the sum of the events

A1 , A1c B1c A2 , A1c B1c A2c B2c A3 , ,


A1c B1c Anc Bnc An+1 ,

and therefore
2.6* The
2.4Probability
Combinatorial
of Matchings
Results 51

(
P A occurs before B occurs )
[ (
= P A1 + A1c B1c A2 + A1c B1c A2c B2c A3 ) ( )
(
+ + A1c B1c Anc Bnc An+1 + ) ]
( ) (
= P A1 + P A1c B1c A2 + P A1c B1c A2c B2c A3 ) ( )
( )
+ + P A1c B1c Anc Bnc An+1 +

= P( A ) + P( A B ) P( A ) + P( A B ) P( A B ) P( A )
1 1
c
1
c
2 1
c
1
c c
2
c
2 3

+ + P( A B ) P( A B )P( A ) + (by Theorem 6)


1
c
1
c c
n
c
n n+1

= P( A) + P( A B )P( A) + P ( A B )P( A)
c c 2 c c

+ + P ( A B )P( A)
n c c

= P( A)[1 + P( A B ) + P ( A B ) + + P ( A B ) + ]
c c 2 c c n c c

( ) 1 P A1 B
=P A .
( c c
)
But

( )
P Ac Bc = P A B = 1 P A B

c


( ) ( )
= 1 P A+ B = 1 P A P B , ( ) ( ) ( )
so that

(
1 P Ac Bc = P A + P B . ) ( ) ()
Therefore

( )
P A
(
P A occurs before B occurs = ) ,
( ) ( )
P A +P B
as asserted.
It is possible to interpret B as a catastrophic event, and A as an event
consisting of taking certain precautionary and protective actions upon the
energizing of a signaling device. Then the significance of the above probability
becomes apparent. As a concrete illustration, consider the following simple
example (see also Exercise 2.6.3).
52 2 Some Probabilistic Concepts and Results

EXAMPLE 11 In repeated (independent) draws with replacement from a standard deck of 52


playing cards, calculate the probability that an ace occurs before a picture.

Let A = an ace occurs, B = a picture occurs.

Then P(A) = 4
52
= 1
13
and P(B) = 12
52
= 4
13
, so that P(A occurs before B occurs)
= 131 134 = 14 .

Exercises
2.6.1 Show that
m + k M M M m
= ,
m m + k m k
as asserted in relation (8).
2.6.2 Verify the transition in (7) and that the resulting expression is indeed
the desired result.
2.6.3 Consider the following game of chance. Two fair dice are rolled repeat-
edly and independently. If the sum of the outcomes is either 7 or 11, the player
wins immediately, while if the sum is either 2 or 3 or 12, the player loses
immediately. If the sum is either 4 or 5 or 6 or 8 or 9 or 10, the player continues
rolling the dice until either the same sum appears before a sum of 7 appears in
which case he wins, or until a sum of 7 appears before the original sum appears
in which case the player loses. It is assumed that the game terminates the first
time the player wins or loses. What is the probability of winning?
3.1 Soem General Concepts 53

Chapter 3

On Random Variables and


Their Distributions

3.1 Some General Concepts


Given a probability space (S, class of events, P), the main objective of prob-
ability theory is that of calculating probabilities of events which may be of
importance to us. Such calculations are facilitated by a transformation of the
sample space S, which may be quite an abstract set, into a subset of the real
line ! with which we are already familiar. This is, actually, achieved by the
introduction of the concept of a random variable. A random variable (r.v.) is
a function (in the usual sense of the word), which assigns to each sample point
s S a real number, the value of the r.v. at s. We require that an r.v. be a well-
behaving function. This is satisfied by stipulating that r.v.s are measurable
functions. For the precise definition of this concept and related results, the
interested reader is referred to Section 3.5 below. Most functions as just
defined, which occur in practice are, indeed, r.v.s, and we leave the matter to
rest here. The notation X(S) will be used for the set of values of the r.v. X, the
range of X.
Random variables are denoted by the last letters of the alphabet X, Y, Z,
etc., with or without subscripts. For a subset B of !, we usually denote by
(X B) the following event in S: (X B) = {s S; X(s) B} for simplicity. In
particular, (X = x) = {s S; X(s) = x}. The probability distribution function (or
just the distribution) of an r.v. X is usually denoted by PX and is a probability
function defined on subsets of ! as follows: PX (B) = P(X B). An r.v. X is said
to be of the discrete type (or just discrete) if there are countable (that is, finitely
many or denumerably infinite) many points in !, x1, x2, . . . , such that PX({xj})
> 0, j 1, and j PX({xj})(= j P(X = xj)) = 1. Then the function fX defined on the
entire ! by the relationships:

( ) ({ })(= P(X = x ))
fX x j = PX x j j for x = x j ,

53
54 3 On Random Variables and Their Distributions

and fX(x) = 0 otherwise has the properties:

()
fX x 0 for all x, and f (x ) = 1.
j
X j

Furthermore, it is clear that

( ) f (x ).
P X B =
x j B
X j

Thus, instead of striving to calculate the probability of the event {s S;


X(s) B}, all we have to do is to sum up the values of fX(xj) for all those xjs
which lie in B; this assumes, of course, that the function fX is known. The
function fX is called the probability density function (p.d.f.) of X. The distribu-
tion of a discrete r.v. will also be referred to as a discrete distribution. In the
following section, we will see some discrete r.v.s (distributions) often occur-
ring in practice. They are the Binomial, Poisson, Hypergeometric, Negative
Binomial, and the (discrete) Uniform distributions.
Next, suppose that X is an r.v. which takes values in a (finite or infinite but
proper) interval I in ! with the following qualification: P(X = x) = 0 for every
single x in I. Such an r.v. is called an r.v. of the continuous type (or just a
continuous r.v.). Also, it often happens for such an r.v. to have a function fX
satisfying the properties fX(x) 0 for all x I, and P(X J) = J fX(x)dx for any
sub-interval J of I. Such a function is called the probability density function
(p.d.f.) of X in analogy with the discrete case. It is to be noted, however, that
here fX(x) does not represent the P(X = x)! A continuous r.v. X with a p.d.f. fX
is called absolutely continuous to differentiate it from those continuous r.v.s
which do not have a p.d.f. In this book, however, we are not going to concern
ourselves with non-absolutely continuous r.v.s. Accordingly, the term con-
tinuous r.v. will be used instead of absolutely continuous r.v. Thus, the r.v.s
to be considered will be either discrete or continuous (= absolutely continu-
ous). Roughly speaking, the idea that P(X = x) = 0 for all x for a continuous r.v.
may be interpreted that X takes on too many values for each one of them to
occur with positive probability. The fact that P(X = x) also follows formally by
the fact that P(X = x) = xx fX(y)dy, and this is 0. Other interpretations are also
possible. It is true, nevertheless, that X takes values in as small a neighborhood
of x as we please with positive probability. The distribution of a continuous r.v.
is also referred to as a continuous distribution. In Section 3.3, we will discuss
some continuous r.v.s (distributions) which occur often in practice. They are
the Normal, Gamma, Chi-square, Negative Exponential, Uniform, Beta,
Cauchy, and Lognormal distributions. Reference will also be made to t and F
r.v.s (distributions).
Often one is given a function f and is asked whether f is a p.d.f. (of some
r.v.). All one has to do is to check whether f is non-negative for all values of its
argument, and whether the sum or integral of its values (over the appropriate
set) is equal to 1.
3.2 Discrete Random Variables
3.1 Soem
(and General
RandomConcepts
Vectors) 55

When (a well-behaving) function X is defined on a sample space S and


takes values in the plane or the three-dimensional space or, more generally, in
the k-dimensional space ! k, it is called a k-dimensional random vector (r.
vector) and is denoted by X. Thus, an r.v. is a one-dimensional r. vector. The
distribution of X, PX, is defined as in the one-dimensional case by simply
replacing B with subsets of ! k. The r. vector X is discrete if P(X = xj) > 0, j =
1, 2, . . . with j P(X = xj) = 1, and the function fX(x) = P(X = xj) for x = xj, and
fX(x) = 0 otherwise is the p.d.f. of X. Once again, P(X B) = xjB fX(xj) for B
subsets of !k. The r. vector X is (absolutely) continuous if P(X = x) = 0 for all
x I, but there is a function fX defined on !k such that:

()
fX x 0 (
for all x ! k , and P X J = fX x dx ) J
()
for any sub-rectangle J of I. The function fX is the p.d.f. of X. The distribution
of a k-dimensional r. vector is also referred to as a k-dimensional discrete or
(absolutely) continuous distribution, respectively, for a discrete or (abso-
lutely) continuous r. vector. In Sections 3.2 and 3.3, we will discuss two repre-
sentative multidimensional distributions; namely, the Multinomial (discrete)
distribution, and the (continuous) Bivariate Normal distribution.
We will write f rather than fX when no confusion is possible. Again, when
one is presented with a function f and is asked whether f is a p.d.f. (of some r.
vector), all one has to check is non-negativity of f, and that the sum of its values
or its integral (over the appropriate space) is equal to 1.

3.2 Discrete Random Variables (and Random Vectors)


3.2.1 Binomial
The Binomial distribution is associated with a Binomial experiment; that is, an
experiment which results in two possible outcomes, one usually termed as a
success, S, and the other called a failure, F. The respective probabilities
are p and q. It is to be noted, however, that the experiment does not really
have to result in two outcomes only. Once some of the possible outcomes are
called a failure, any experiment can be reduced to a Binomial experiment.
Here, if X is the r.v. denoting the number of successes in n binomial experi-
ments, then
n
( ) {
X S = 0, 1, 2, , n , } ( ) ()
P X = x = f x = p x q n x ,
x
where 0 < p < 1, q = 1 p, and x = 0, 1, 2, . . . , n. That this is in fact a p.d.f.
follows from the fact that f(x) 0 and
n n
n
f ( x ) = x p x q n x = ( p + q)
n
= 1n = 1.
x =0 x =0
56 3 On Random Variables and Their Distributions

The appropriate S here is:

{ }
S = S, F S, F{ } (n copies).
In particular, for n = 1, we have the Bernoulli or Point Binomial r.v. The r.v.
X may be interpreted as representing the number of Ss (successes) in
the compound experiment E E (n copies), where E is the experiment
resulting in the sample space {S, F} and the n experiments are independent
(or, as we say, the n trials are independent). f(x) is the probability that exactly
x Ss occur. In fact, f(x) = P(X = x) = P(of all n sequences of Ss and Fs
with exactly x Ss). The probability of one such a sequence is pxqnx by the
independence of the trials and this also does not depend on the particular
sequence we are considering. Since there are ( xn) such sequences, the result
follows.
The distribution of X is called the Binomial distribution and the quantities
n and p are called the parameters of the Binomial distribution. We denote the
Binomial distribution by B(n, p). Often the notation X B(n, p) will be used
to denote the fact that the r.v. X is distributed as B(n, p). Graphs of the p.d.f.
of the B(n, p) distribution for selected values of n and p are given in Figs. 3.1
and 3.2.

f (x)
0.25
n ! 12
0.20 1
p! 4
0.15
0.10
0.05
x
0 1 2 3 4 5 6 7 8 9 10 11 12 13
Figure 3.1 Graph of the p.d.f. of the Binomial distribution for n = 12, p = 14 .

()
f 0 = 0.0317 ()
f 7 = 0.0115
f (1) = 0.1267 f ( 8 ) = 0.0024
f (2 ) = 0.2323 f (9) = 0.0004
f ( 3) = 0.2581 f (10 ) = 0.0000
f ( 4 ) = 0.1936 f (11) = 0.0000
f ( 5) = 0.1032 f (12 ) = 0.0000
f (6) = 0.0401
3.2 3.1 Soem
Discrete Random Variables (and General
RandomConcepts
Vectors) 57

f (x)
0.25

0.20 n ! 10
1
0.15 p! 2

0.10
0.05
x
0 1 2 3 4 5 6 7 8 9 10
Figure 3.2 Graph of the p.d.f. of the Binomial distribution for n = 10, p = 12 .

()
f 0 = 0.0010 ()
f 6 = 0.2051
f (1) = 0.0097 f ( 7) = 0.1172
f (2 ) = 0.0440 f ( 8 ) = 0.0440
f ( 3) = 0.1172 f (9) = 0.0097
f ( 4 ) = 0.2051 f (10 ) = 0.0010
f ( 5) = 0.2460

3.2.2 Poisson
x
( ) {
X S = 0, 1, 2, , } ( ) ()
P X = x = f x = e
x!
,

x = 0, 1, 2, . . . ; > 0. f is, in fact, a p.d.f., since f(x) 0 and


x
f ( x ) = e x! = e

e = 1.
x =0 x =0

The distribution of X is called the Poisson distribution and is denoted by


P(). is called the parameter of the distribution. Often the notation X
P() will be used to denote the fact that the r.v. X is distributed as P().
The Poisson distribution is appropriate for predicting the number of
phone calls arriving at a given telephone exchange within a certain period
of time, the number of particles emitted by a radioactive source within a
certain period of time, etc. The reader who is interested in the applications
of the Poisson distribution should see W. Feller, An Introduction to
Probability Theory, Vol. I, 3rd ed., 1968, Chapter 6, pages 156164, for further
examples.
In Theorem 1 in Section 3.4, it is shown that the Poisson distribution
58 3 On Random Variables and Their Distributions

may be taken as the limit of Binomial distributions. Roughly speaking, sup-


pose that X B(n, p), xwhere n is large and p is small. Then P X = x = ( )
( xn ) px (1 p) e np ( npx !) , x 0 . For the graph of the p.d.f. of the P()
n x

distribution for = 5 see Fig. 3.3.


A visualization of such an approximation may be conceived by stipulating
that certain events occur in a time interval [0,t] in the following manner: events
occurring in nonoverlapping subintervals are independent; the probability
that one event occurs in a small interval is approximately proportional to its
length; and two or more events occur in such an interval with probability
approximately 0. Then dividing [0,t] into a large number n of small intervals of
length t/n, we have that the probability that exactly x events occur in [0,t] is
approximately ( nx )( nt ) (1 nt ) , where is the factor of proportionality.
x n x

Setting pn = nt , we havex npn = t and Theorem 1 in Section 3.4 gives that


( nx )( nt ) (1 nt ) e t (xt )! . Thus Binomial probabilities are approximated by
x n x

Poisson probabilities.

f (x)

0.20

0.15

0.10

0.05
x
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Figure 3.3 Graph of the p.d.f. of the Poisson distribution with = 5.

()
f 0 = 0.0067 ()
f 9 = 0.0363
f (1) = 0.0337 f (10 ) = 0.0181
f (2 ) = 0.0843 f (11) = 0.0082
f ( 3) = 0.1403 f (12 ) = 0.0035
f ( 4 ) = 0.1755 f (13) = 0.0013
f ( 5) = 0.1755 f (14 ) = 0.0005
f (6) = 0.1462 f (15) = 0.0001
f ( 7) = 0.1044
f ( 8 ) = 0.0653 ()
f n is negligible for n 16.
3.2 Discrete Random Variables
3.1 Soem
(and General
RandomConcepts
Vectors) 59

3.2.3 Hypergeometric
m n

x r x
( ) {
X S = 0, 1, 2, , r , } ()
f x =
m + n
,

r

where ( mx) = 0, by definition, for x > m. f is a p.d.f., since f(x) 0 and


r r
m n m + n
f ( x ) = m + n x r x = m + n
1 1
= 1.
x =0 x =0 r

r r
The distribution of X is called the Hypergeometric distribution and arises in
situations like the following. From an urn containing m red balls and n black
balls, r balls are drawn at random without replacement. Then X represents the
number of red balls among the r balls selected, and f(x) is the probability that
this number is exactly x. Here S = {all r-sequences of Rs and Bs}, where R
stands for a red ball and B stands for a black ball. The urn/balls model just
described is a generic model for situations often occurring in practice. For
instance, the urn and the balls may be replaced by a box containing certain
items manufactured by a certain process over a specified period of time, out of
which m are defective and n meet set specifications.

3.2.4 Negative Binomial


r + x 1 x
( ) {
X S = 0, 1, 2, , } ()
f x = pr
x
q ,

0 < p < 1, q = 1 p, x = 0, 1, 2, . . . . f is, in fact, a p.d.f. since f(x) 0 and



r + x 1 x pr pr
f ( x ) = p
r
q = = = 1.
x ( )
r
x =0 x =0 1 q pr

This follows by the Binomial theorem, according to which

1
n + j 1 j
= x , x < 1.
(1 x) j=0
n
j

The distribution of X is called the Negative Binomial distribution. This distri-


bution occurs in situations which have as a model the following. A Binomial
experiment E, with sample space {S, F}, is repeated independently until exactly
r Ss appear and then it is terminated. Then the r.v. X represents the number
of times beyond r that the experiment is required to be carried out, and f(x) is
the probability that this number of times is equal to x. In fact, here S =
60 3 On Random Variables and Their Distributions

{all (r + x)-sequences of Ss and Fs such that the rth S is at the end of the
sequence}, x = 0, 1, . . . and f(x) = P(X = x) = P[all (r + x)-sequences as above
for a specified x]. The probability of one such sequence is pr1qxp by the
independence assumption, and hence
r + x 1 r 1 x r r + x 1 x
()
f x =
x
p q p= p
x
q .

The above interpretation also justifies the name of the distribution. For r = 1,
we get the Geometric (or Pascal) distribution, namely f(x) = pqx, x = 0, 1, 2, . . . .

3.2.5 Discrete Uniform

( ) {
X S = 0, 1, , n 1 , } ()
f x =
1
n
, x = 0, 1, , n 1.

This is the uniform probability measure. (See Fig. 3.4.)

f (x)
n!5
1 Figure 3.4 Graph of the p.d.f. of a Discrete
5 Uniform distribution.
x
0 1 2 3 4

3.2.6 Multinomial
Here
k
() ( )
X S = x = x1 , , xk ; x j 0, j = 1, 2, , k, x j = n,
j =1
k

()
f x =
n!
x1! x 2 ! x k !
p1x 1 p2x 2 pkx k , p j > 0, j = 1, 2, , k, p j = 1.
j =1

That f is, in fact, a p.d.f. follows from the fact that

f ( x) = ( )
n! n
p1x1 pkxk = p1 + + pk = 1n = 1,
X x 1 , , xk x1! xk !
where the summation extends over all xjs such that xj 0, j = 1, 2, . . . , k,
kj=1 xj = n. The distribution of X is also called the Multinomial distribution and
n, p1, . . . , pk are called the parameters of the distribution. This distribution
occurs in situations like the following. A Multinomial experiment E with k
possible outcomes Oj, j = 1, 2, . . . , k, and hence with sample space S = {all
n-sequences of Ojs}, is carried out n independent times. The probability of
the Ojs occurring is pj, j = 1, 2, . . . k with pj > 0 and kj=1 p j = 1 . Then X is the
random vector whose jth component Xj represents the number of times xj
the outcome Oj occurs, j = 1, 2, . . . , k. By setting x = (x1, . . . , xk), then f is the
3.1 Soem General Exercises
Concepts 61

probability that the outcome Oj occurs exactly xj times. In fact f(x) = P(X = x)
= P(all n-sequences which contain exactly xj Ojs, j = 1, 2, . . . , k). The prob-
ability of each one of these sequences is p1x1 pkxk by independence, and since
there are n!/( x1! xk ! ) such sequences, the result follows.
The fact that the r. vector X has the Multinomial distribution with param-
eters n and p1, . . . , pk may be denoted thus: X M(n; p1, . . . , pk).
REMARK 1 When the tables given in the appendices are not directly usable
because the underlying parameters are not included there, we often resort to
linear interpolation. As an illustration, suppose X B(25, 0.3) and we wish
to calculate P(X = 10). The value p = 0.3 is not included in the Binomial
Tables in Appendix III. However, 164 = 0.25 < 0.3 < 0.3125 = 165 and the
probabilities P(X = 10), for p = 164 and p = 165 are, respectively, 0.9703
and 0.8756. Therefore linear interpolation produces the value:

0.3 0.25
(
0.9703 0.9703 0.8756 ) 0.3125 0.25
= 0.8945.

Likewise for other discrete distributions. The same principle also applies
appropriately to continuous distributions.
REMARK 2 In discrete distributions, we are often faced with calculations of
the form x =1 x x . Under appropriate conditions, we may apply the following
approach:

d x d
d
x x
= x x 1 = = x
= = .
d d d 1
( )
2
x =1 x =1 x =1 x =1 1

Similarly for the expression


x=2 (
x x1 ) x2
.

Exercises
3.2.1 A fair coin is tossed independently four times, and let X be the r.v.
defined on the usual sample space S for this experiment as follows:

()
X s = the number of H s in s.

iii) What is the set of values of X?


iii) What is the distribution of X?
iii) What is the partition of S induced by X?
3.2.2 It has been observed that 12.5% of the applicants fail in a certain
screening test. If X stands for the number of those out of 25 applicants who fail
to pass the test, what is the probability that:
62 3 On Random Variables and Their Distributions

iii) X 1?
iii) X 20?
iii) 5 X 20?
3.2.3 A manufacturing process produces certain articles such that the prob-
ability of each article being defective is p. What is the minimum number, n, of
articles to be produced, so that at least one of them is defective with probabil-
ity at least 0.95? Take p = 0.05.

3.2.4 If the r.v. X is distributed as B(n, p) with p > 12 , the Binomial Tables
in Appendix III cannot be used directly. In such a case, show that:
ii) P(X = x) = P(Y = n x), where Y B(n, q), x = 0, 1, . . . , n, and q = 1 p;
ii) Also, for any integers a, b with 0 a < b n, one has: P(a X b) =
P(n b Y n a), where Y is as in part (i).
3.2.5 Let X be a Poisson distributed r.v. with parameter . Given that
P(X = 0) = 0.1, compute the probability that X > 5.
3.2.6 Refer to Exercise 3.2.5 and suppose that P(X = 1) = P(X = 2). What is
the probability that X < 10? If P(X = 1) = 0.1 and P(X = 2) = 0.2, calculate the
probability that X = 0.
3.2.7 It has been observed that the number of particles emitted by a radio-
active substance which reach a given portion of space during time t follows
closely the Poisson distribution with parameter . Calculate the probability
that:
iii) No particles reach the portion of space under consideration during
time t;
iii) Exactly 120 particles do so;
iii) At least 50 particles do so;
iv) Give the numerical values in (i)(iii) if = 100.
3.2.8 The phone calls arriving at a given telephone exchange within one
minute follow the Poisson distribution with parameter = 10. What is the
probability that in a given minute:
iii) No calls arrive?
iii) Exactly 10 calls arrive?
iii) At least 10 calls arrive?
3.2.9 (Truncation of a Poisson r.v.) Let the r.v. X be distributed as Poisson
with parameter and define the r.v. Y as follows:

(
Y = X if X k a given positive integer ) and Y = 0 otherwise.

Find:
3.1 Soem General Exercises
Concepts 63

ii) P(Y = y), y = k, k + 1, . . . ;


ii) P(Y = 0).
3.2.10 A university dormitory system houses 1,600 students, of whom 1,200
are undergraduates and the remaining are graduate students. From the com-
bined list of their names, 25 names are chosen at random. If X stands for the
r.v. denoting the number of graduate students among the 25 chosen, what is
the probability that X 10?
3.2.11 (Multiple Hypergeometric distribution) For j = 1, . . . , k, consider an
urn containing nj balls with the number j written on them. n balls are drawn at
random and without replacement, and let Xj be the r.v. denoting the number
of balls among the n ones with the number j written on them. Then show that
the joint distribution of Xj, j = 1, . . . , k is given by

(
P X j = x j , j = 1, , k = ) (

k
( ),
nj
j =1 x j
n1 + + nk
n )
k
0 x j n j , j = 1, , k, x j = n.
j =1

3.2.12 Refer to the manufacturing process of Exercise 3.2.3 and let Y be the
r.v. denoting the minimum number of articles to be manufactured until the
first two defective articles appear.
ii) Show that the distribution of Y is given by

( ) ( )( )
y2
P Y = y = p2 y 1 1 p , y = 2, 3, ;

ii) Calculate the probability P(Y 100) for p = 0.05.


3.2.13 Show that the function f(x) = ( 12 )xIA(x), where A = {1, 2, . . .}, is a p.d.f.
3.2.14 For what value of c is the function f defined below a p.d.f.?

() ()
f x = c x I A x , where {
A = 0, 1, 2, } (0 < < 1).
3.2.15 Suppose that the r.v. X takes on the values 0, 1, . . . with the following
probabilities:

() (
f j =P X=j = ) c
3j
, j = 0, 1, ;

iii) Determine the constant c.


Compute the following probabilities:
iii) P(X 10);
iii) P(X A), where A = {j; j = 2k + 1, k = 0, 1, . . .};
iv) P(X B), where B = {j; j = 3k + 1, k = 0, 1, . . .}.
64 3 On Random Variables and Their Distributions

3.2.16 There are four distinct types of human blood denoted by O, A, B and
AB. Suppose that these types occur with the following frequencies: 0.45, 0.40,
0.10, 0.05, respectively. If 20 people are chosen at random, what is the prob-
ability that:
ii) All 20 people have blood of the same type?
ii) Nine people have blood type O, eight of type A, two of type B and one of
type AB?
3.2.17 A balanced die is tossed (independently) 21 times and let Xj be the
number of times the number j appears, j = 1, . . . , 6.
ii) What is the joint p.d.f. of the Xs?
ii) Compute the probability that X1 = 6, X2 = 5, X3 = 4, X4 = 3, X5 = 2, X6 = 1.
3.2.18 Suppose that three coins are tossed (independently) n times and
define the r.v.s Xj, j = 0, 1, 2, 3 as follows:

X j = the number of times j H s appear.

Determine the joint p.d.f. of the Xjs.


3.2.19 Let X be an r.v. distributed as P(), and set E = {0, 2, . . . } and O = {1,
3, . . . }. Then:
ii) In terms of , calculate the probabilities: P(X E) and P(X O);
ii) Find the numerical values of the probabilities in part (i) for = 5. (Hint: If
k k
SE = kE k! and SO = kO k! , notice that SE + SO = e, and SE SO =

e .)
3.2.20 The following recursive formulas may be used for calculating
Binomial, Poisson and Hypergeometric probabilities. To this effect, show
that:
iii) If X B(n, p), then f ( x + 1) = qp nx+x1 f ( x), x = 0, 1, . . . , n 1;
iii) If X P(), then f ( x + 1) = x+1 f ( x), x = 0, 1, . . . ;
iii) If X has the Hypergeometric distribution, then

(m x)(r x) f x ,
(
f x+1 = )
(n r + x + 1)(x + 1)
() { }
x = 0, 1, , min m, r .

3.2.21
i) Suppose the r.v.s X1, . . . , Xk have the Multinomial distribution, and let
j be a fixed number from the set {1, . . . , k}. Then show that Xj is distributed as
B(n, pj);
ii) If m is an integer such that 2 m k 1 and j1, . . . , jm are m distinct integers
from the set {1, . . . , k}, show that the r.v.s Xj , . . . , Xj have Multinomial
1 m

distributions with parameters n and pj , pj .


1 m
3.3 Continuous Random Variables
3.1 Soem
(and General
RandomConcepts
Vectors) 65

3.2.22 (Polyas urn scheme) Consider an urn containing b black balls and r
red balls. One ball is drawn at random, is replaced and c balls of the same color
as the one drawn are placed into the urn. Suppose that this experiment is
repeated independently n times and let X be the r.v. denoting the number of
black balls drawn. Then show that the p.d.f. of X is given by

( )(
b b + c b + 2c b + x 1 c [ ( )]
)
n r (r + c ) [r + ( n x 1)c ]
(
P X =x = ) .
x b + r b + r + c ( )( )
( )
b + r + 2c b + r + m 1 c [ ( )]
(This distribution can be used for a rough description of the spread of conta-
gious diseases. For more about this and also for a certain approximation to the
above distribution, the reader is referred to the book An Introduction to
Probability Theory and Its Applications, Vol. 1, 3rd ed., 1968, by W. Feller, pp.
120121 and p. 142.)

3.3 Continuous Random Variables (and Random Vectors)


3.3.1 Normal (or Gaussian)
x
( )
2

( )
X S = !, f x = () 1
exp
2 2
,

x ! .
2
We say that X is distributed as normal (, 2), denoted by N(, 2), where ,
2 are called the parameters of the distribution of X which is also called
the Normal distribution ( = mean, !, 2 = variance, > 0). For = 0,
= 1, we get what is known as the Standard Normal distribution, denoted

by N(0, 1). Clearly f(x) > 0; that I = f x dx = 1 is proved by showing that ()
I = 1. In fact,
2

()
I 2 = f x dx = f x dx f y dy () ()

x 2
( y
) ( )
2
1 1
dx exp dy
2
= exp
2 2 2 2

1 1 1

2 2
= e z 2
dz e 2
d ,
2

upon letting (x )/ = z, so that z (, ), and (y )/ = , so that


(, ). Thus
66 3 On Random Variables and Their Distributions

f (x)
0.8
$ ! 0.5
0.6

0.4
$!1
0.2
$!2
x
"2 "1 0 1 2 3 4 5
N(#, $ 2 )
Figure 3.5 Graph of the p.d.f. of the Normal distribution with = 1.5 and several values of .

1 (
z 2 + 2 ) 2 dz d = 1 2
2
2
I2 = e e r 2
r dr d
2 0 0

by the standard transformation to polar coordinates. Or

1 2
r dr d = e r
2 2 2
I2 = e r 2 2
r dr = e r 2
0 = 1;
2 0 0 0

that is, I2 = 1 and hence I = 1, since f(x) > 0.


It is easily seen that f(x) is symmetric about x = , that is, f( x) =
f( + x) and that f(x) attains its maximum at x = which is equal to 1/( 2 ).
From the fact that

max f x = () 1
x !
2
and the fact that

f ( x)dx = 1,

we conclude that the larger is, the more spread-out f(x) is and vice versa. It
is also clear that f(x) 0 as x . Taking all these facts into consideration,
we are led to Fig. 3.5.
The Normal distribution is a good approximation to the distribution of
grades, heights or weights of a (large) group of individuals, lifetimes of various
manufactured items, the diameters of hail hitting the ground during a storm,
the force required to punctuate a cardboard, errors in numerous measure-
ments, etc. However, the main significance of it derives from the Central Limit
Theorem to be discussed in Chapter 8, Section 8.3.

3.3.2 Gamma
( ) (
X S = ! actually X S = 0, ( ) ( ))
3.3 Continuous Random Variables
3.1 Soem
(and General
RandomConcepts
Vectors) 67

Here
1
x 1 e x , x > 0 > 0, > 0,
()
f x = ( )
0 , x0


where () = 0 y 1 e y dy (which exists and is finite for > 0). (This integral
is known as the Gamma function.) The distribution of X is also called the
Gamma distribution and , are called the parameters of the distribution.

Clearly, f(x) 0 and that f(x)dx = 1 is seen as follows.

f (x )dx = ( )
1 1

( )
x 1 e x dx = y 1 e y dy,
0
0

upon letting x/ = y, x = y, dx = dy, y (0, ); that is,

f (x )dx = ( ) 0 ( )
1 1
y 1 e y dy = = 1.
( )

REMARK 3 One easily sees, by integrating by parts, that

( ) (
= 1 1 , )( )
and if is an integer, then

( ) ( )(
= 1 2 1 , ) ()
where

() () ( )

1 = e y dy = 1; that is, a = 1 !
0

We often use this notation even if is not an integer, that is, we write

( ) ( )

= 1 ! = y 1 e y dy for > 0.
0

For later use, we show that

1 1
= ! = .
2 2
We have
1 (1 2 )
= y e y dy.

2 0

By setting

t2
y1 2 =
t
, so that y=
2
, dy = t dt , t 0, .( )
2
68 3 On Random Variables and Their Distributions

we get

1 1
= 2 e t t dt = 2 e t
2 2
2 2
dt = ;

2 0 t 0

that is,

1 1
= ! = .
2 2
From this we also get that

3 1 1
= = , etc.

2 2
2 2
Graphs of the p.d.f. of the Gamma distribution for selected values of and
are given in Figs. 3.6 and 3.7.
The Gamma distribution, and in particular its special case the Negative
Exponential distribution, discussed below, serve as satisfactory models for

f (x)
1.00

% ! 1, & ! 1
0.75

0.50
% ! 2, & ! 1
0.25 % ! 4, & ! 1

x
0 1 2 3 4 5 6 7 8
Figure 3.6 Graphs of the p.d.f. of the Gamma distribution for several values of , .

f (x)

1.00

0.75 % ! 2, & ! 0.5

0.50
% ! 2, & ! 1
0.25
% ! 2, & ! 2
x
0 1 2 3 4 5
Figure 3.7 Graphs of the p.d.f. of the Gamma distribution for several values of , .
3.3 Continuous Random Variables
3.1 Soem
(and General
RandomConcepts
Vectors) 69

describing lifetimes of various manufactured items, among other things, as


well as in statistics.
For specific choices of the parameters and in the Gamma distribution,
we obtain the Chi-square and the Negative Exponential distributions given
below.

3.3.3 Chi-square
For = r/2, r 1, integer, = 2, we get what is known as the Chi-square
distribution, that is,

1 ( r 2 )1 x 2
x e , x>0
() 1
( )
f x = 2 r 2

r 2
r > 0, integer.
0, x0
The distribution with this p.d.f. is denoted by r2 and r is called the number of
degrees of freedom (d.f.) of the distribution. The Chi-square distribution occurs
often in statistics, as will be seen in subsequent chapters.

3.3.4 Negative Exponential


For = 1, = 1/, we get

e x , x>0
()
f x =
x0
> 0,
0 ,
which is known as the Negative Exponential distribution. The Negative Expo-
nential distribution occurs frequently in statistics and, in particular, in waiting-
time problems. More specifically, if X is an r.v. denoting the waiting time
between successive occurrences of events following a Poisson distribution,
then X has the Negative Exponential distribution. To see this, suppose that
events occur according to the Poisson distribution P(); for example, particles
emitted by a radioactive source with the average of particles per time unit.
Furthermore, we suppose that we have just observed such a particle, and let X
be the r.v. denoting the waiting time until the next particle occurs. We shall
show that X has the Negative Exponential distribution with parameter .
To this end, it is mentioned here that the distribution function F of an r.v., to
be studied in the next chapter, is defined by F(x) = P(X x), x !, and if X
is a continuous r.v., then dFdx( x ) = f ( x). Thus, it suffices to determine F here.
Since X 0, it follows that F(x) = 0, x 0. So let x > 0 be the the waiting time
for the emission of the next item. Then F(x) = P(X x) = 1 P(X > x). Since
is the average number of emitted particles per time unit, their average
x ( x )
0

number during time x will be x. Then P( X > x) = e 0!


= e x , since no
particles are emitted in (0, x]. That is, F(x) = 1 e , so that f(x) = ex. To
x

summarize: f(x) = 0 for x 0, and f(x) = ex for x > 0, so that X is distributed


as asserted.
70 3 On Random Variables and Their Distributions

Consonant with previous notation, we may use the notation X (, ) or


X NE(), or X r2 in order to denote the fact that X is distributed as
Gamma with parameters and , or Negative Exponental with parameter ,
or Chi-square with r degrees of freedom, respectively.

3.3.5 Uniform U(, ) or Rectangular R(, )


( ) (
X S = ! actually X S = , and ( ) [ ])
()
f x = (
1 , ) x
< .
0, otherwise
Clearly,

() f (x )dx = dx = 1.
1

f x 0,

The distribution of X is also called Uniform or Rectangular (, ), and and


are the parameters of the distribution. The interpretation of this distribution
is that subintervals of [, ], of the same length, are assigned the same prob-
ability of being observed regardless of their location. (See Fig. 3.8.)
The fact that the r.v. X has the Uniform distribution with parameters
and may be denoted by X U(, ).

f (x)

1
&"% Figure 3.8 Graph of the p.d.f. of the
U(, ) distribution.

x
0 % &

3.3.6 Beta
( ) (
X S = ! actually X S = 0, 1 and ( ) ( ))
+ ( )
( )
1
x 1 1 x 0<x<1
()
f x = ( )()
,

0 elsewhere, > 0, > 0.

Clearly, f(x) 0. That f ( x)dx = 1 is seen as follows.
3.3 Continuous Random Variables
3.1 Soem
(and General
RandomConcepts
Vectors) 71

( )()
= x 1 e x dx y 1 e y dy

0 0
( x+y )
= x 1 y 1 e dx dy
0 0

which, upon setting u = x/(x + y), so that

x=
uy
, dx =
y du
, u 0, 1 ( ) and x + y =
y
,
1 u
( ) 1 u
2
1 u
becomes

1 u 1 (
y 1 u ) du
= y 1 y 1 e y dy
(1 u) (1 u)
0 0 1 2

1 u 1 y 1 u( )
= y + 1 e du dy.
( )
0 0 +1
1 u

Let y/(1 u) = , so that y = (1 u), dy = (1 u)d, (0, ). Then the


integral is

u (1 u)
1 1
= 1
+ 1 e du d
0 0

( )
1 1
= + 1 e d u 1 1 u du
0 0

( ) u (1 u)
1
1 1
= + du;
0

that is,

( )() ( ) x (1 x)
1
1 1
= + dx
0

and hence

(
+ )
f (x )dx = ( )( ) 0 x (1 x )
1 1
1
dx = 1.

Graphs of the p.d.f. of the Beta distribution for selected values of and are
given in Fig. 3.9.
REMARK 4 For = = 1, we get the U(0, 1), since (1) = 1 and (2) = 1.
The distribution of X is also called the Beta distribution and occurs rather
often in statistics. , are called the parameters of the distribution
and the function defined by 0 x 1 (1 x ) 1 dx for , > 0 is called the Beta
1

function.
Again the fact that X has the Beta distribution with parameters and
may be expressed by writing X B(, ).
72 3 On Random Variables and Their Distributions

f (x) %!5
&!3
2.5
%!3
&!3
2.0

%!2
1.5
&!2

1.0

x
0 0.2 0.4 0.6 0.8 1.0
Figure 3.9 Graphs of the p.d.f. of the Beta distribution for several values of , .

3.3.7 Cauchy
Here


( )
X S = ! and ()
f x =
1
, x ! , ! , > 0.
2 + x
( )
2

Clearly, f(x) > 0 and


f ( x )dx =
1 1 1
1 + x
dx = dx
( ) [( ) ]
2 2
2 + x
1 dy 1

1 + y 2
= = arctan y = 1,

upon letting

x dx
y= , so that = dy.

The distribution of X is also called the Cauchy distribution and , are called
the parameters of the distribution (see Fig. 3.10). We may write X
Cauchy(, 2) to express the fact that X has the Cauchy distribution with
parameters and 2.
(The p.d.f. of the Cauchy distribution looks much the same as the Normal
p.d.f., except that the tails of the former are heavier.)

3.3.8 Lognormal
Here X(S) = ! (actually X(S) = (0, )) and
3.3 3.1 Soem
Continuous Random Variables (and General
RandomConcepts
Vectors) 73

f (x)

0.3

0.2

0.1

x
"2 "1 0 1 2
Figure 3.10 Graph of the p.d.f. of the Cauchy distribution with = 0, = 1.


( )
2
1 exp log x log , x>0

()
f x = x 2


2 2


0, x 0 where , > 0.
Now f(x) 0 and
log x log
( )
2

()
1 1
f x dx = 0 exp dx

2 x 2 2

which, letting x = ey, so that log x = y, dx = eydy, y (, ), becomes
y log
( )
2
1 1

= y 2 2
2 e
exp e y dy.


But this is the integral of an N(log , 2) density and hence is equal to 1; that
is, if X is lognormally distributed, then Y = log X is normally distributed with
parameters log and 2. The distribution of X is called Lognormal and , are
called the parameters of the distribution (see Fig. 3.11). The notation X
Lognormal(, ) may be used to express the fact that X has the Lognormal
distribution with parameters and .
(For the many applications of the Lognormal distribution, the reader is
referred to the book The Lognormal Distribution by J. Aitchison and J. A. C.
Brown, Cambridge University Press, New York, 1957.)

"
These distributions occur very often in Statistics (interval esti-
3.3.9 t
mation, testing hypotheses, analysis of variance, etc.) and their
3.3.10 F densities will be presented later (see Chapter 9, Section 9.2).
We close this section with an example of a continuous random vector.

3.3.11 Bivariate Normal


Here X(S) = !2 (that is, X is a 2-dimensional random vector) with
74 3 On Random Variables and Their Distributions

f (x)
0.8 & 2 ! 0.5
%!1

0.6

1
0.4 % ! e2

0.2 %!e

x
0 1 2 3 4
Figure 3.11 Graphs of the p.d.f. of the Lognormal distribution for several values of , .

(
f x1 , x 2 = ) 1
e q 2 ,
2 1 2 1 2

where x1, x2 !; 1, 2 > 0; 1 < < 1 and

1 x 2 x 1 1 x 2 2 x 2 2
2

q= 1 1
2 +

1 2 1 1 2 2

with 1, 2 !. The distribution of X is also called the Bivariate Normal
distribution and the quantities 1, 2, 1, 2, are called the parameters of the
distribution. (See Fig. 3.12.)
Clearly, f(x1, x2) > 0. That ! f(x1, x2)dx1dx2 = 1 is seen as follows:
2

x 2 x1 1 x 2 2 x 2 2
2

( )
1 2 q = 1
1
1
2
1 2 2
+


2 2
x 2 x1 1
= 2
2

1 1
(
2 x1 1
+ 1 . )
Furthermore,

x2 2 x1 1 x 2 2 1 x 1
= 2 1
2 1 2 2 1
x1 1
=
1
2
x 2 2 + 2

=

1
x2 b , ( )
1 2

where

2
b = 2 +
1
(
x1 1 . )
3.3 3.1 Soem
Continuous Random Variables (and General
RandomConcepts
Vectors) 75

z
z ! f(x1, x 2 ) for z ' k

x2
0

x1 k

Figure 3.12 Graph of the p.d.f. of the Bivariate Normal distribution.

Thus
2 2

(
1 2 q = 2)
x b
2
2 x1 1
+ 1
1
( )
and hence

x
(
)
2

( )
1
exp
1 1
f x1 , x 2 dx 2 =
2 1 2 12


( )
2
1 x2 b dx
exp

2 2 1 2
2 2 1 2
2 ( ) 2

x
( )
2
1
exp 1,
1 1
=
2 1 2 2
1

since the integral above is that of an N(b, 22(1 2)) density. Since the first
factor is the density of an N(1, 12) random variable, integrating with respect
to x1, we get

f ( x1 , x 2 )dx1 dx 2 = 1.

REMARK 5 From the above derivations, it follows that, if f(x1, x2) is Bivariate
Normal, then

( ) ( ) ( )

f1 x1 = f x1 , x 2 dx 2 is N 1 , 12 ,

and similarly,

( ) ( ) ( )

f2 x 2 = f x1 , x 2 dx1 is N 2 , 22 .

76 3 On Random Variables and Their Distributions

As will be seen in Chapter 4, the p.d.f.s f1 and f2 above are called marginal
p.d.f.s of f.
The notation X N(1, 2, 12, 22, ) may be used to express the fact that
X has the Bivariate Normal distribution with parameters 1, 2, 12, 22, .
Then X1 N(1 , 12) and X2 N(2, 22 ).

Exercises
3.3.1 Let f be the p.d.f. of the N(, 2) distribution and show that:
ii) f is symmetric about ;

ii) max f x = () 1
.
x !
2
3.3.2 Let X be distributed as N(0, 1), and for a < b, let p = P(a < X < b). Then
use the symmetry of the p.d.f. f in order to show that:
iii) For 0 a < b, p = (b) (a);
iii) For a 0 < b, p = (b) + (a) 1;
iii) For a b < 0, p = (a) (b);
iv) For c > 0, P(c < X < c) = 2(c) 1.
(See Normal Tables in Appendix III for the definition of .)
3.3.3 If X N(0, 1), use the Normal Tables in Appendix III in order to show
that:
iii) P(1 < X < 1) = 0.68269;
iii) P(2 < X < 2) = 0.9545;
iii) P(3 < X < 3) = 0.9973.
3.3.4 Let X be a r2 . In Table 5, Appendix III, the values = P(X x) are
given for r ranging from 1 to 45, and for selected values of . From the entries
of the table, observe that, for a fixed , the values of x increase along with the
number of degrees of freedom r. Select some values of and record the
corresponding values of x for a set of increasing values of r.
3.3.5 Let X be an r.v. distributed as 102. Use Table 5 in Appendix III in order
to determine the numbers a and b for which the following are true:
ii) P(X < a) = P(X > b);
ii) P(a < X < b) = 0.90.
3.3.6 Consider certain events which in every time interval [t1, t2] (0 < t1 < t2)
occur independently for nonoverlapping intervals according to the Poisson
distribution P((t2 t1)). Let T be the r.v. denoting the time which lapses
3.1 Soem General Exercises
Concepts 77

between two consecutive such events. Show that the distribution of T is Nega-
tive Exponential with parameter by computing the probability that T > t.
3.3.7 Let X be an r.v. denoting the life length of a TV tube and suppose that
its p.d.f. f is given by:
f(x) = exI(0, )(x).
Compute the following probabilities:
iii)P(j < X j + 1), j = 0, 1, . . . ;
iii)P(X > t) for some t > 0;
iii)P(X > s + t|X > s) for some s, t > 0;
iv) Compare the probabilities in parts (ii) and (iii) and conclude that the
Negative Exponential distribution is memoryless;
iv) If it is known that P(X > s) = , express the parameter in terms of
and s.
3.3.8 Suppose that the life expectancy X of each member of a certain group
of people is an r.v. having the Negative Exponential distribution with param-
eter = 1/50 (years). For an individual from the group in question, compute
the probability that:
iii) He will survive to retire at 65;
iii) He will live to be at least 70 years old, given that he just celebrated his 40th
birthday;
iii) For what value of c, P(X > c) = 12 ?
3.3.9 Let X be an r.v. distributed as U(, ) ( > 0). Determine the values
of the parameter for which the following are true:
ii) P(1 < X < 2) = 0.75;
ii) P(|X| < 1) = P(|X| > 2).
3.3.10 Refer to the Beta distribution and set:
1
( ) ( )
1
B , = x 1 1 x dx.
0

Then show that B(, ) = B(, ).


3.3.11 Establish the following identity:

n 1 p m1
( ) n!
( )
n m p n m
n x 1 x dx = x m1 1 x dx
m 1 0 ( )(
m1 nm ! 0 )
n

= n p j 1 p ( )
n j
.
j= m
j

3.3.12 Let X be an r.v. with p.d.f given by f(x) = 1/[(1 + x2)]. Calculate the
probability that X2 c.
78 3 On Random Variables and Their Distributions

3.3.13 Show that the following functions are p.d.f.s:


= xex /2I(0, )(x) (Raleigh distribution);
2
iii) f(x)
= 2 / x2ex /2I(0,)(x) (Maxwells distribution);
2
iii) f(x)
iii) f(x) = 12 e|x| (Double Exponential);
iv) f(x) = ( ca )( cx )+1IA(x), A = (c, ), , c > 0 (Pareto distribution).

3.3.14 Show that the following functions are p.d.f.s:


ii) f(x) = cos xI(0,/2)(x);
ii) f(x) = xexI(0,)(x).
3.3.15 For what values of the constant c are the following functions p.d.f.s?

ce 6 x , x>0

ii) f(x) = cx, 1 < x 0;
0 , x 1

ii) f(x) = cx2ex I(0,) (x).
3

3.3.16 Let X be an r.v. with p.d.f. given by 3.3.15(ii). Compute the probabil-
ity that X > x.
3.3.17 Let X be the r.v. denoting the life length of a certain electronic device
expressed in hours, and suppose that its p.d.f. f is given by:

()
f x =
c
()
I 1,000 , 3,000 ] x .
xn [

ii) Determine the constant c in terms of n;


ii) Calculate the probability that the life span of one electronic device of the
type just described is at least 2,000 hours.
3.3.18 Refer to Exercise 3.3.15(ii) and compute the probability that X ex-
ceeds s + t, given that X > s. Compare the answer with that of Exercise
3.3.7(iii).
3.3.19 Consider the function f(x) = x 1e x , x > 0 (, > 0), and:

iii) Show that it is a p.d.f. (called the Weibull p.d.f. with parameters
and );
iii) Observe that the Negative Exponential p.d.f. is a special case of a Weibull
p.d.f., and specify the values of the parameters for which this happens;
iii) For = 1 and = 12 , = 1 and = 2, draw the respective graphs of the
p.d.f.s involved.
(Note: The Weibull distribution is employed for describing the lifetime of
living organisms or of mechanical systems.)
3.4
3.1 The
SoemPoisson
General
Distribution
Concepts 79

3.3.20 Let X and Y be r.v.s having the joint p.d.f. f given by:

( ) ( ) (
f x, y = c 25 x 2 y 2 I (0 ,5 ) x 2 + y 2 . )
Determine the constant c and compute the probability that 0 < X2 + Y2 < 4.
3.3.21 Let X and Y be r.v.s whose joint p.d.f. f is given by f(x, y) =
cxyI(0,2)(0,5)(x, y). Determine the constant c and compute the following
probabilities:
i) P( 12 < X < 1, 0 < Y < 3);
ii) P(X < 2, 2 < Y < 4);
iii) P(1 < X < 2, Y > 5);
iv) P(X > Y).
3.3.22 Verify that the following function is a p.d.f.:

( )
f x, y =
1
4
( ) (
cos y I A x, y , ) ( ]
A = , , .
2 2
3.3.23 (A mixed distribution) Show that the following function is a p.d.f.
1 x
e , x0
4
1 , 0<x<2

()
f x = 8
x
1
, x = 2, 3,
2
0, otherwise.

3.4 The Poisson Distribution as an Approximation to


the Binomial Distribution and the Binomial Distribution as
an Approximation to the Hypergeometric Distribution
In this section, we first establish rigorously the assertion made earlier that the
Poisson distribution may be obtained as the limit of Binomial distributions. To
this end, consider a sequence of Binomial distributions, so that the nth distri-
bution has probability pn of a success, and we assume that as n , pn 0 and
that
n = npn ,
for some > 0. Then the following theorem is true.
THEOREM 1 With the above notation and assumptions, we have


x
n x n x
pn q n n
e for each fixed x = 0, 1, 2, .
x
x!
80 3 On Random Variables and Their Distributions

PROOF We have

( ) (
n x n x n n 1 n x + 1 x n x
pn q n = pn q n
)
x x!

=
( ) ( )
n n 1 n x + 1 n n
x n x

1
x! n n

=
( ) (
n n1 nx +1 ) 1 x
n

1 n
1
n
nx x! n n
x

1
n
1 x 1 nx
= 1 1 1
n n x!
n
1 x
1 n

n
e ,
n n
x
x!
1
n

since, if n , then
n
n
1 e .
n

This is merely a generalization of the better-known fact that


n

1 e .
n

REMARK 6 The meaning of the theorem is the following: If n is large the


probabilities ( nx )pxqnx are not easily calculated. Then we can approximate
them by e(x/x!), provided p is small, where we replace be np. This is true
for all x = 0, 1, 2, . . . , n.
We also meet with difficulty in calculating probabilities in the Hyper-
geometric distribution
m n m + n

x r x r
if m, n are large. What we do is approximate the Hypergeometric distribution
by an appropriate Binomial distribution, and then, if need be, we can go one
step further in approximating the Binomial by the appropriate Poisson distri-
bution according to Theorem 1. Thus we have
THEOREM 2 Let m, n and suppose that m/(m + n) = pm,n p, 0 < p < 1. Then
3.4
3.1 The
SoemPoisson
General
Distribution
Concepts 81

m n

x r x r
p x q r x , x = 0, 1, 2, , r .
m + n x

r

PROOF We have

m n

x r x
=
(
m! n! m + n r ! )
r!
m + n

( )[ ( )] ( )
m x ! n r x ! m + n ! x! r x ! ( )
r
( ) (
r m x + 1 m n r + x + 1 n
=
)
x (
m+nr +1 m+n) ( )
r m( m 1) [ m ( x 1)] n( n 1) [ n (r x 1)]
= .
x (m + n) [(m + n) (r 1)]
Both numerator and denominator have r factors. Dividing through by (m + n),
we get

m x

n r x r m m 1 m x1
=
m + n x m + n m + n m + n m + n m + n

r
n n 1 n r x 1

m + n m + n m + n m+n m+n
1
1 r 1
1 1 1
m + n m + n
r
m ,n
p x q r x ,
x

since

m n
p and hence 1 p = q. "
m+n m+n

REMARK 7 The meaning of the theorem is that if m, n are large, we can


approximate the probabilities
82 3 On Random Variables and Their Distributions

m n

x r x r
by p x q r x
m + n x

r
by setting p = m/(m +x n). This is true for all x = 0, 1, 2, . . . , r. It is to be
observed that ( xr )( mm+ n ) (1 mm+ n ) is the exact probability of having exactly x
rx

successes in r trials when sampling is done with replacement, so that the


probability of a success, mm+ n , remains constant. The Hypergeometric distribu-
tion is appropriate when sampling is done without replacement. If, however,
m
m and n are large (m, n ) and n remains approximately constant
( mn c = p / q), then the probabilities of having exactly x successes in r
(independent) trials are approximately equal under both sampling schemes.

Exercises
3.4.1 For the following values of n, p and = np, draw graphs of B(n, p) and
P() on the same coordinate axes:
iii) n= 10, p = 4
16
, so that = 2.5;
iii) n= 16, p = 2
16
, so that = 2;
iii) n= 20, p = 2
16
, so that = 2.5;
iv) n= 24, p = 1
16
, so that = 1.5;
iv) n= 24, p = 2
16
, so that = 3.
3.4.2 Refer to Exercise 3.2.2 and suppose that the number of applicants is
equal to 72. Compute the probabilities (i)(iii) by using the Poisson approxi-
mation to Binomial (Theorem 1).
3.4.3 Refer to Exercise 3.2.10 and use Theorem 2 in order to obtain an
approximate value of the required probability.

3.5* Random Variables as Measurable Functions and Related Results


In this section, random variables and random vectors are introduced as special
cases of measurable functions. Certain results related to -fields are also
derived. Consider the probability space (S, A, P) and let T be a space and X
be a function defined on S into T, that is, X : S T. For T T, define the
inverse image of T, under X, denoted by X1(T), as follows:

( ) { () }
X 1 T = s S ; X s T .
3.5* Random Variables as Measurable Functions
3.1 Soem and
General
Related
Concepts
Results 83

This set is also denoted by [X T] or (X T). Then the following


properties are immediate consequences of the definition (and the fact X is a
function):

j j
( )
X 1 U T j = U X 1 T j . (1)

( ) ( )
If T1 T2 = , then X 1 T1 X 1 T2 = . (2)
Hence by (1) and (2) we have

j j
( )
X 1 T j = X 1 T j . (3)


Also
j j
( )
X 1 I T j = I X 1 T j , (4)

( ) [ ( )] ,
c
X 1 T c = X 1 T (5)

( )
X 1 T = S , (6)

X 1 () = . (7)

Let now D be a -field of subsets of T and define the class X1(D) of


subsets of S as follows:

( ) { ( )
X 1 D = A S ; A = X 1 T for some T D . }
By means of (1), (5), (6) above, we immediately have
THEOREM 3 The class X1(D) is a -field of subsets of S.
The above theorem is the reason we require measurability in our defini-
tion of a random variable. It guarantees that the probability distribution
function of a random vector, to be defined below, is well defined.
If X1(D) A, then we say that X is (A, D)-measurable, or just measur-
able if there is no confusion possible. If (T, D) = (!, B) and X is (A, B)-
measurable, we say that X is a random variable (r.v.). More generally, if (T, D)
= (!k, Bk), where !k = ! ! ! (k copies of !), and X is (A, Bk)-
measurable, we say that X is a k-dimensional random vector (r. vector). In this
latter case, we shall write X if k 1, and just X if k = 1. A random variable is
a one-dimensional random vector.
On the basis of the properties (1)(7) of X1, the following is immediate.
THEOREM 4 Define the class C* of subsets of T as follows: C* = {T T; X1(T) = A for some
A A}. Then C* is a -field.
COROLLARY Let D = (C), where C is a class of subsets of T. Then X is (A, D)-measurable
if and only if X1(C) A. In particular, X is a random variable if and only if
84 3 On Random Variables and Their Distributions

X1(Co), or X1(Cj), or X1(C j) A, j = 1, 2, . . . , 8, and similarly for the case


of k-dimensional random vectors. The classes Co, Cj, C j, j = 1, . . . , 8 are
defined in Theorem 5 and the paragraph before it in Chapter 1.
PROOF The -field C* of Theorem 4 has the property that C* C. Then C*
D = (C) and hence X1(C*) X1(D). But X1(C*) A. Thus X1(D)
A. The converse is a direct consequence of the definition of (A, D) =
measurability.
Now, by means of an r. vector X: (S, A, P) (!k, Bk ), define on Bk the
set function PX as follows:

( ) [ ( )] ( ) ({
PX B = P X 1 B = P X B = P s S ; X s B . ( ) }) (8)
By the Corollary to Theorem 4, the sets X1(B) in S are actually events due to
the assumption that X is an r. vector. Therefore PX is well defined by (8); i.e.,
P[X1(B)] makes sense, is well defined. It is now shown that PX is a probability
measure on Bk . In fact, PX(B) 0, B Bk, since P is a probability measure.
Next, PX(!k) = P[X 1(!k)] = P(S ) = 1, and finally,


[ ( )]

( ) ( )
PX B j = P X 1 B j = P X 1 B j = P X 1 B j = PX B j .
j =1 j = 1
j =1 j =1 j =1

The probability measure PX is called the probability distribution function (or


just the distribution) of X.

Exercises
3.5.1 Consider the sample space S supplied with the -field of events A. For
an event A, the indicator IA of A is defined by: IA(s) = 1 if s A and IA(s) = 0
if s Ac.
iii) Show that IA is r.v. for any A A .
iii) What is the partition of S induced by IA?
iii) What is the -field induced by IA?
3.5.2 Write out the proof of Theorem 1 by using (1), (5) and (6).
3.5.3 Write out the proof of Theorem 2.
4.1 The Cumulative Distribution Function 85

Chapter 4

Distribution Functions, Probability


Densities, and Their Relationship

4.1 The Cumulative Distribution Function (c.d.f. or d.f.) of a Random Vector


Basic Properties of the d.f. of a Random Variable
The distribution of a k-dimensional r. vector X has been defined through the
relationship: PX (B) = P(X B), where B is a subset of ! k. In particular, one
may choose B to be an interval in ! k; i.e., B = {y ! k; y x} in the sense that,
if x = (x1, . . . , xk) and y = (y1, . . . , yk), then yj xj, j = 1, . . . , k. For such a choice
of B, PX(B) is denoted by FX(x) and is called the cumulative distribution
function of X (evaluated at x), or just the distribution function (d.f.) of X. We
omit the subscript X if no confusion is possible. Thus, the d.f. F of a r. vector
X is an ordinary point function defined on ! k (and taking values in [0, 1]). Now
we restrict our attention to the case k = 1 and prove the following basic
properties of the d.f. of an r.v.
THEOREM 1 The distribution function F of a random variable X satisfies the following
properties:
i) 0 F(x) 1, x !.
ii) F is nondecreasing.
iii) F is continuous from the right.
iv) F(x) 0 as x , F(x) 1, as x +.
We express this by writing F() = 0, F(+) = 1.
PROOF In the course of this proof, we set Q for the distribution PX of X, for
the sake of simplicity. We have then:
i) Obvious.
ii) This means that x1 < x2 implies F(x1) F(x2). In fact,

85
86 4 Distribution Functions, Probability Densities, and Their Relationship

x1 < x2 implies (, x ] (, x ]
1 2

and hence
( ] ( ]
Q , x1 Q , x 2 ; equivalently, F x1 F x 2 . ( ) ( )
iii) This means that, if xn x, then F(xn) F(x). In fact,
x n x implies (, x ] (, x]
n

and hence
( ]
Q , x n Q , x ( ]
by Theorem 2, Chapter 2; equivalently, F(xn) F(x).
iv) Let xn . We may assume that xn (see also Exercise 4.1.6). Then
(, x ] ,
n (
so that Q , x n Q = 0] ( )
by Theorem 2, Chapter 2. Equivalently, F(xn) 0. Similarly, if xn +.
We may assume xn . Then
(, x ] ! and hence Q(, x ] Q(! ) = 1; equivalently, F ( x ) 1. "
n n n

Graphs of d.f.s of several distributions are given in Fig. 4.1.


REMARK 1
i) F(x) can be used to find probabilities of the form P(a < X b); that is
(
P a<X b =F b F a . ) () ()
In fact,
(a < X b) = ( < X b) ( < X a)
and
( < X a) ( < X b).
Thus
( ) ( ) (
P a < X b = P < X b P < X a = F b F a . ) () ()
ii) The limit from the left of F(x) at x, denoted by F(x), is defined as follows:
( )
F x = lim F x n
n
( ) with x n x.

This limit always exists, since F(xn), but need not be equal to F(x+)(=limit
from the right) = F(x). The quantities F(x) and F(x) are used to express
the probability P(X = a); that is, P(X = a) = F(a) F(a). In fact, let xn
a and set A = (X = a), An = (xn < X a). Then, clearly, An A and hence
by Theorem 2, Chapter 2,
( ) ( )
P An P A , or lim P x n < X a = P X = a ,
n
( ) ( )
or

[ ( ) ( )] (
lim F a F x n = P X = a ,
n
)
4.1 The Cumulative Distribution Function 87

or
() ( )
F a lim F x n = P X = a ,
n
( )
or
() ( )
F a F a =P X =a . ( )
It is known that a nondecreasing function (such as F) may have
discontinuities which can only be jumps. Then F(a) F(a) is the length of
the jump of F at a. Of course, if F is continuous then F(x) = F(x) and
hence P(X = x) = 0 for all x.
iii) If X is discrete, its d.f. is a step function, the value of it at x being defined
by
( ) f (x )
F x =
xjx
j and ( ) ( ) ( )
f x j = F x j F x j1 ,

where it is assumed that x1 < x2 < .


iv) If X is of the continuous type, its d.f. F is continuous. Furthermore,
dF x ( )=f
dx
(x)

F(x) F(x)
1.00 1.00
0.80 0.80
0.60 0.60
0.40 0.40
0.20 0.20
0 x 0 x
1
(a) Binomial for n ! 6, p ! 4
. (b) Poisson for ( ! 2.

F(x) )(x)

1.0 1.0

0 x 0.5
% &
x
0 x+% "2 "1 0 1 2
x"%
(c) U (%, &). Here F(x) ! &"%
% * x * &.
(d) N(0, 1).
1 x'&
Figure 4.1 Examples of graphs of c.d.f.s.
88 4 Distribution Functions, Probability Densities, and Their Relationship

at continuity points of f, as is well known from calculus. Through the


relations
dF x ( )=f
() () (x),
x
F x = f t dt and
dx
we see that if f is continuous, f determines F( f F) and F determines f (F
f ); that is, F f. Two important applications of this are the following two
theorems.
Often one has to deal with functions of an r. vector itself. In such cases, the
resulting entities have got to be r. vectors, since we operate in a probability
framework. The following statement is to this effect. Its precise formulation
and justification is given as Theorem 7 on page 104.
STATEMENT 1 Let X be a k-dimensional r. vector defined on the sample space S, and let g be
a (well-behaving) function defined on ! k and taking values in ! m. Then g(X)
is defined on the underlying sample space S, takes values in ! m, and is an r.
vector. (That is, well-behaving functions of r. vectors are r. vectors.) In particu-
lar, g(X) is an r. vector if g is continuous.
Now a k-dimensional r. vector X may be represented in terms of its
coordinates; i.e., X = (X1, . . . , Xk), where Xj, j = 1, . . . , k are real-valued
functions defined on S. The question then arises as to how X and Xj, j = 1, . . . ,
k are related from the point of view of being r. vectors. The answer is provided
by the following statement, whereas the precise statement and justification are
given as Theorem 8 below.
STATEMENT 2 Let X and Xj, j = 1, . . . , k be functions defined on the sample space S and
taking values in ! k and !, respectively, and let X = (X1, . . . , Xk). Then X is an
r. vector if and only if Xj, j = 1, . . . , k are r.v.s.
The following two theorems provide applications of Statement 1.
THEOREM 2 Let X be an N(, 2)-distributed r.v. and set Y = X
. Then Y is an r.v. and its
distribution is N(0, 1).
PROOF In the first place, Y is an r.v. by Statement 1. Then it suffices to show
that the d.f. of Y is , where

() 1 y

2
y = e t 2
dt .

2
We have

X
( )
P Y y = P

y = P X y +

( )
(
t
)
2
1 1
()
y +
exp dt = y
e
u2 2
= du = y ,
2 2 2 2

4.1 The Cumulative DistributionExercises
Function 89

where we let u = (t )/ in the transformation of the integral. "


REMARK 2 The transformation X

of X is referred to as normalization
of X.
THEOREM 3 i(i) Let X be an N(0, 1)-distributed r.v. Then Y = X 2 is distributed as 21.
(ii) If X is a N ( , 2)-distributed r.v., then the r.v. ( X )2 is distributed as 21.
PROOF (i) We will show that the p.d.f. of Y is that of a 21-distributed r.v. by
deriving the d.f. of Y first and then differentating it in order to obtain fY. To this
end, let us observe first that Y is an r.v. on account of Statement 1. Next, for
y > 0, we have

() ( )
FY y = P Y y = P y X ( y )
1 y 1 y

2 2
= ex 2
dx = 2 ex 2
dx.
y
2 2 0

Let x = t . Then dx = dt/2 t , t (0, y] and

() 1 1
y
FY y = 2
2

0
2 t
e t 2 dt .

Hence
dFY y( )= 1 1
e y 2
=
1
y
(1 2 )1
e y 2 .
dy 2 y 2

Since fY(y) = 0 for y 0 (because FY(y) = 0, y 0), it follows that

1 (1 2 )1 y 2
y e , y>0
()
fY y = 21 2
0, y 0,

and this is the p.d.f. of 21. (Observe that here we used the fact that ( 12 ) =
.) "

Exercises
4.1.1 Refer to Exercise 3.2.13, in Chapter 3, and determine the d.f.s corre-
sponding to the p.d.f.s given there.
4.1.2 Refer to Exercise 3.2.14, in Chapter 3, and determine the d.f.s corre-
sponding to the p.d.f.s given there.
90 4 Distribution Functions, Probability Densities, and Their Relationship

4.1.3 Refer to Exercise 3.3.13, in Chapter 3, and determine the d.f.s corre-
sponding to the p.d.f.s given there.
4.1.4 Refer to Exercise 3.3.14, in Chapter 3, and determine the d.f.s corre-
sponding to the p.d.f.s given there.
4.1.5 Let X be an r.v. with d.f. F. Determine the d.f. of the following r.v.s:
X, X 2, aX + b, XI[a,b) (X ) when:
i) X is continuous and F is strictly increasing;
ii) X is discrete.
4.1.6 Refer to the proof of Theorem 1 (iv) and show that we may assume
that xn (xn ) instead of xn (xn ).
4.1.7 Let f and F be the p.d.f. and the d.f., respectively, of an r.v. X. Then
show that F is continuous, and dF(x)/dx = f(x) at the continuity points x of f.
4.1.8
i) Show that the following function F is a d.f. (Logistic distribution) and
derive the corresponding p.d.f., f.

()
F x =
1
(x + )
, x ! , > 0, ! ;
1+ e
ii) Show that f(x) =F(x)[1 F(x)].
4.1.9 Refer to Exercise 3.3.17 in Chapter 3 and determine the d.f. F corre-
sponding to the p.d.f. f given there. Write out the expressions of F and f for
n = 2 and n = 3.
4.1.10 If X is an r.v. distributed as N(3, 0.25), use Table 3 in Appendix III in
order to compute the following probabilities:
i) P(X < 1);
ii) P(X > 2.5);
iii) P(0.5 < X < 1.3).
4.1.11 The distribution of IQs of the people in a given group is well approxi-
mated by the Normal distribution with = 105 and = 20. What proportion of
the individuals in the group in question has an IQ:
i) At least 150?
ii) At most 80?
iii) Between 95 and 125?
4.1.12 A certain manufacturing process produces light bulbs whose life
length (in hours) is an r.v. X distributed as N(2,000, 2002). A light bulb is
supposed to be defective if its lifetime is less than 1,800. If 25 light bulbs are
4.2 The 4.1
d.f. ofThe
a Random
Cumulative
Vector
Distribution
and Its Properties
Function 91

tested, what is the probability that at most 15 of them are defective? (Use the
required independence.)
4.1.13 A manufacturing process produces 12 -inch ball bearings, which are
assumed to be satisfactory if their diameter lies in the interval 0.5 0.0006 and
defective otherwise. A days production is examined, and it is found that the
distribution of the actual diameters of the ball bearings is approximately
normal with mean = 0.5007 inch and = 0.0005 inch. Compute the propor-
tion of defective ball bearings.
4.1.14 If X is an r.v. distributed as N(, 2), find the value of c (in terms of
and ) for which P(X < c) = 2 9P(X > c).
4.1.15 Refer to the Weibull p.d.f., f, given in Exercise 3.3.19 in Chapter 3 and
do the following:
i) Calculate the corresponding d.f. F and the reliability function !(x) = 1
F(x);
()
f (x )
ii) Also, calculate the failure (or hazard) rate H x = ! ( x ) , and draw its graph
for = 1 and = 1 , 1, 2;
2
iii) For s and t > 0, calculate the probability P(X > s + t|X > t) where X is an r.v.
having the Weibull distribution;
iv) What do the quantities F(x), !(x), H(x) and the probability in
part (iii) become in the special case of the Negative Exponential
distribution?

4.2 The d.f. of a Random Vector and Its PropertiesMarginal and Conditional
d.f.s and p.d.f.s
For the case of a two-dimensional r. vector, a result analogous to Theorem
1 can be established. So consider the case that k = 2. We then have X =
(X1, X2) and the d.f. F(or FX or FX1,X2) of X, or the joint distribution function
of X1, X2, is F(x1, x2) = P(X1 x1, X2 x2). Then the following theorem holds
true.
With the above notation we have
THEOREM 4
i) 0 F(x1, x2) 1, x1, x2 !.
ii) The variation of F over rectangles with sides parallel to the axes, given in
Fig. 4.2, is 0.
iii) F is continuous from the right with respect to each of the coordinates x1, x2,
or both of them jointly.
92 4 Distribution Functions, Probability Densities, and Their Relationship

y
(x1, y 2)
y2 (x2, y 2)
" ,
Figure 4.2 The variation V of F over the
rectangle is:
, "
y1 (x2, y 1) F(x1, y1) + F(x2, y2) F(x1, y2) F(x2, y1)
(x1, y 1)
x
0 x1 x2

iv) If both x1, x2, , then F(x1, x2) 1, and if at least one of the x1, x2
, then F(x1, x2) 0. We express this by writing F(, ) = 1, F(, x2) =
F(x1, ) = F(, ) = 0, where < x1, x2 < .
PROOF
i) Obvious.
ii) V = P(x1 < X1 x2, y1 < X2 y2) and is hence, clearly, 0.
iii) Same as in Theorem 3. (If x = (x1, x2), and zn = (x1n, x2n), then zn x means
x1n x1, x2n x2).
iv) If x1, x2 , then (, x1] (, x2] R 2, so that F(x1, x2) P(S) = 1. If
at least one of x1, x2 goes () to , then (, x1] (, x2] , hence
( ) ( )
F x1 , x 2 P = 0. "

REMARK 3 The function F(x1, ) = F1(x1) is the d.f. of the random variable
X1. In fact, F(x1, ) = F1(x1) is the d.f. of the random variable X1. In fact,

( ) (
F x1 , = lim P X1 x1 , X 2 x n
x n
)
( ) (
= P X1 x1 , < X 2 < = P X1 x1 = F1 x1 . ) ( )
Similarly F(, x2) = F2(x2) is the d.f. of the random variable X2. F1, F2 are called
marginal d.f.s.
REMARK 4 It should be pointed out here that results like those discussed
in parts (i)(iv) in Remark 1 still hold true here (appropriately interpreted).
In particular, part (iv) says that F(x1, x2) has second order partial derivatives
and
2
x1x 2
( ) (
F x1 , x 2 = f x1 , x 2 )
at continuity points of f.
For k > 2, we have a theorem strictly analogous to Theorems 3 and 6 and
also remarks such as Remark 1(i)(iv) following Theorem 3. In particular, the
analog of (iv) says that F(x1, . . . , xk) has kth order partial derivatives and
4.2 The 4.1
d.f. ofThe
a Random
Cumulative
Vector
Distribution
and Its Properties
Function 93

k
x1x2 xk
(
F x1 , , xk = f x1 , , xk ) ( )
at continuity points of f, where F, or FX, or FX1, , Xk, is the d.f. of X, or the joint
distribution function of X1, . . . , Xk. As in the two-dimensional case,

(
F , , , x j , , , = F j x j ) ( )
is the d.f. of the random variable Xj, and if m xjs are replaced by (1 < m < k),
then the resulting function is the joint d.f. of the random variables correspond-
ing to the remaining (k m) Xjs. All these d.f.s are called marginal distribu-
tion functions.
In Statement 2, we have seen that if X = (X1, . . . , Xk) is an r. vector, then
Xj, j = 1, 2, . . . , k are r.v.s and vice versa. Then the p.d.f. of X, f(x) =
f(x1, . . . , xk), is also called the joint p.d.f. of the r.v.s X1, . . . , Xk.
Consider first the case k = 2; that is, X = (X1, X2), f(x) = f(x1, x2) and set

f x1 , x 2

( )
( )
f1 x1 = x 2

( )

f x1 , x 2 dx 2


f x1 , x 2

( )
( )
f2 x 2 = x 1

( )

f x1 , x 2 dx1 .

Then f1, f2 are p.d.f.s. In fact, f1(x1) 0 and

f ( x ) = f ( x , x ) = 1,
1 1 1 2
x1 x1 x2

or

f1 (x1 )dx1 = f (x1 , x 2 )dx1 dx 2 = 1.


Similarly we get the result for f2. Furthermore, f1 is the p.d.f. of X1, and f2 is the
p.d.f. of X2. In fact,

(
f x1 , x2 = f x1 , x2 = f1 x1

) ( ) ( )
( )
P X 1 B = x B , x !
1 2 x B x ! x B 1 2 1

( )
f x1 , x2 dx1dx2 = f x1 , x2 dx2 dx1 = f1 x1 dx1 .
B ! B ! B [ ( ) ] ( )
Similarly f2 is the p.d.f. of the r.v. X2. We call f1, f2 the marginal p.d.f.s. Now
suppose f1(x1) > 0. Then define f(x2|x1) as follows:

) (f (x ) ) .
f x,x
(
f x 2 x1 =
1
1

1
2
94 4 Distribution Functions, Probability Densities, and Their Relationship

This is considered as a function of x2, x1 being an arbitrary, but fixed, value of


X1 ( f1(x1) > 0). Then f(|x1) is a p.d.f. In fact, f(x2|x1) 0 and

f ( x 2 x1 ) = f ( x1 , x 2 ) = ( )
1 1
f1 x1 = 1,
x2 ( )
f1 x1 x2 ( )
f1 x1

f ( x 2 x1 )dx 2 = f ( x ) f ( x1 , x 2 )dx 2 = f ( x ) f1 ( x1 ) = 1.
1 1
1 1 1 1

In a similar fashion, if f2(x2) > 0, we define f(x1|x2) by:

) (f (x ) )
f x, x
(
f x1 x 2 =
2
1

2
2

and show that f(|x2) is a p.d.f. Furthermore, if X1, X2 are both discrete, the
f(x2|x1) has the following interpretation:

) (f (x ) ) = ( P(X = x ) ) = P(X
f x,x P X =x, X =x
(
f x 2 x1 =
1
1

1
2 1 1

1
2

1
2
2 = x 2 X1 = x1 . )
Hence P(X2 B|X1 = x1) = x B f(x2|x1). For this reason, we call f(|x2) the
2

conditional p.d.f. of X2, given that X1 = x1 (provided f1(x1) > 0). For a similar
reason, we call f(|x2) the conditional p.d.f. of X1, given that X2 = x2 (provided
f2(x2) > 0). For the case that the p.d.f.s f and f2 are of the continuous type,
the conditional p.d.f. f (x1|x2) may be given an interpretation similar to the
one given above. By assuming (without loss of generality) that h1, h2 > 0, one
has

(1 h )P(x < X x + h x < X x + h )


1 1 1 1 1 2 2 2 2

=
(1 h h )P(x < X x + h , x < X x + h )
1 2 1 1 1 1 2 2 2 2

(1 h )P(x < X x + h )
2 2 2 2 2

=
(1 h h )[F (x , x ) + F (x + h , x + h ) F (x , x + h ) F (x
1 2 1 2 1 1 2 2 1 2 2 1 + h1 , x 2 )]
(1 h )[F (x + h ) F (x )] 2 2 2 2 2 2

where F is the joint d.f. of X1, X2 and F2 is the d.f. of X2. By letting h1, h2 0
and assuming that (x1, x2) and x2 are continuity points of f and f2, respectively,
the last expression on the right-hand side above tends to f(x1, x2)/f2(x2) which
was denoted by f(x1|x2). Thus for small h1, h2, h1 f(x1|x2) is approximately equal
to P(x1 < X1 x1 + h1|x2 < X2 x2 + h2), so that h1 f(x1|x2) is approximately the
conditional probability that X1 lies in a small neighborhood (of length h1) of x1,
given that X2 lies in a small neighborhood of x2. A similar interpretation may
be given to f(x2|x1). We can also define the conditional d.f. of X2, given X1 = x1,
by means of
4.2 The 4.1
d.f. ofThe
a Random
Cumulative
Vector
Distribution
and Its Properties
Function 95

f x x ( )
(
)
2 1
F x2 x1 = x x 2 2

x f x x dx ,
2 1 2
2

( )
and similarly for F(x1|x2).
The concepts introduced thus far generalize in a straightforward way for
k > 2. Thus if X = (X1, . . . , Xk) with p.d.f. f(x1, . . . , xk), then we have called
f(x1, . . . , xk) the joint p.d.f. of the r.v.s X1, X2, . . . , Xk. If we sum (integrate)
over t of the variables x1, . . . , xk keeping the remaining s fixed (t + s = k), the
resulting function is the joint p.d.f. of the r.v.s corresponding to the remaining
s variables; that is,

f x1 , , xk
x , , x
( )
(
fi , , i xi , , xi ) = j1 jt

( )
1 s 1 s
f x ,
1 , xk dx j dx j . 1 t

There are
k k k
+ + + =2 2
k

1 2 k 1
such p.d.f.s which are also called marginal p.d.f.s. Also if xi1, . . . , xis are such
that fi1, . . . , it (xi1, . . . , xis) > 0, then the function (of xj1, . . . , xjt) defined by

(
f x1 , , xk )
(
f x j , , x j xi , , xi =
1 t 1 s ) (
fi , , i xi , , xi )
1 s 1 s

is a p.d.f. called the joint conditional p.d.f. of the r.v.s Xj1, . . . , Xjt, given Xi1 =
xi1, , Xjs = xjs, or just given Xi1, . . . , Xis. Again there are 2k 2 joint condi-
tional p.d.f.s involving all k r.v.s X1, . . . , Xk. Conditional distribution func-
tions are defined in a way similar to the one for k = 2. Thus

(
F x j , , x j xi , , xi
1 t 1 t)

f (x , , x x , , x )
j j i i 1 t 1 s

= ( x , , x )( x , , x
j1 jt j1 jt )
x x f x ,

jt

(
j , x j x i , , x i dx j dx j .
j1

1 t 1 s ) 1 t

We now present two examples of marginal and conditional p.d.f.s, one


taken from a discrete distribution and the other taken from a continuous
distribution.
EXAMPLE 1 Let the r.v.s X1, . . . , Xk have the Multinomial distribution with parameters
n and p1, . . . , pk. Also, let s and t be integers such that 1 s, t < k and
s + t = k. Then in the notation employed above, we have:
96 4 Distribution Functions, Probability Densities, and Their Relationship

ii) 1 s
(
fi , , i xi , , xi =
1 s
) n!
xi ! xi ! n r !
x

(
x
pi pi q n r ,
)
i1

1 s
is

1 s

(
q = 1 pi + + pi , r = xi + + xi ;
1 s
) 1 s

that is, the r.v.s Xi1, . . . , Xis and Y = n (Xi1 + + Xis) have the Multinomial
distribution with parameters n and pi1, . . . , pis, q.

(n r )! p j1
x j1
pj t
xj

ii) (
f x j1 , , x jt xi1 , , xi s ) =
x j1 ! x jt ! q
1 ,
q
r = xi1 + + xi s ;
that is, the (joint) conditional distribution of Xj1, . . . , Xjt given Xi1, . . . , Xis is
Multinomial with parameters n r and pj1/q, . . . , pjt/q.
DISCUSSION
i) Clearly,
(X i1 1 s s
) (
= xi , , X i = xi X i + + X i = r = n Y = r = Y = n r ,
1 s
) ( ) ( )
so that
(X i1 s s
) (
= xi , , X i = xi = X i = xi , , X i = xi , Y = n r .
1 1 1 s s
)
Denoting by O the outcome which is the grouping of all n outcomes distinct
from those designated by i1, . . . , is , we have that the probability of O is q, and
the number of its occurrences is Y. Thus, the r.v.s Xi1, . . . , Xis and Y are
distributed as asserted.
ii) We have

) ( f ( x , , x ) ) = f ((x , , x ))
f x , , x ,x , ,x f x, ,x
(
f x j1 , , x jt xi1 , , xi s =
j1

i1
jt i1

is
is 1

i1
k

is

n! 1 k
n! x x

= p1x pkx pi1i1 pi si s q n r
x1! xk ! xi1 ! xi s ! n r ! ( )

p p p p
xi1 xi s x j1 x jt p p q
xi1 xi s x j1 + + x jt
i is j1 jt i is
= 1 1
xi1 ! xi s ! x j1 ! x jt !



xi ! xi ! n r !
1 s ( )
(since n r = n ( x i1 )
+ + xi s = x j1 + + x jt )
=
(n r )! p j1
x j1
pj
t ,
x jt

x j1 ! x jt ! q q

as was to be seen.
EXAMPLE 2 Let the r.v.s X1 and X2 have the Bivariate Normal distribution, and recall that
their (joint) p.d.f. is given by:
4.1 The Cumulative DistributionExercises
Function 97

(
f x1 , x2 = ) 1
2 1 2 1 2
x 2 2
1 x1 1 x2 2 x2 2
exp 1 1
2 + .
(
2 1
2
) 1


1 2 2

We saw that the marginal p.d.f.s f1, f2 are N(1, 21), N(2, 22), respectively;
that is, X1, X2 are also normally distributed. Furthermore, in the process of
proving that f(x1, x2) is a p.d.f., we rewrote it as follows:


x
( )
( )
2 2
x b
(
f x1 , x 2 = ) 1
exp

1

2 12
1 exp

2
2
,
2 1 2 1 2 2
2 2 1

where

2
b = 2 +
1
(
x1 1 . )
Hence


( )= ( )
2
f x1 , x 2 x2 b
(
f x 2 x1 = ) ( )
f1 x1 2 2
1
1 2
exp
2
2

2 2 1

which is the p.d.f. of an N(b, 22(1 2)) r.v. Similarly f(x1|x2) is seen to be the
p.d.f. of an N(b, 21(1 2)) r.v., where
1
b = 1 +
2
(
x2 2 . )

Exercises
4.2.1 Refer to Exercise 3.2.17 in Chapter 3 and:
i) Find the marginal p.d.f.s of the r.v.s Xj, j = 1, , 6;
ii) Calculate the probability that X1 5.
4.2.2 Refer to Exercise 3.2.18 in Chapter 3 and determine:
ii) The marginal p.d.f. of each one of X1, X2, X3;
ii) The conditional p.d.f. of X1, X2, given X3; X1, X3, given X2; X2, X3, given
X1 ;
98 4 Distribution Functions, Probability Densities, and Their Relationship

iii) The conditional p.d.f. of X1, given X2, X3; X2, given X3, X1; X3, given X1,
X2 .
If n = 20, provide expressions for the following probabilities:
iv) P(3X1 + X2 5);
v) P(X1 < X2 < X3);
vi) P(X1 + X2 = 10|X3 = 5);
vii) P(3 X1 10|X2 = X3);
viii) P(X1 < 3X2|X1 > X3).
4.2.3 Let X, Y be r.v.s jointly distributed with p.d.f. f given by f(x, y) = 2/c2
if 0 x y, 0 y c and 0 otherwise.
i) Determine the constant c;
ii) Find the marginal p.d.f.s of X and Y;
iii) Find the conditional p.d.f. of X, given Y, and the conditional p.d.f. of Y,
given X;
iv) Calculate the probability that X 1.
4.2.4 Let the r.v.s X, Y be jointly distributed with p.d.f. f given by f(x, y) =
exy I(0,)(0,) (x, y). Compute the following probabilities:
i) P(X x);
ii) P(Y y);
iii) P(X < Y);
iv) P(X + Y 3).
4.2.5 If the joint p.d.f. f of the r.v.s Xj, j = 1, 2, 3, is given by

( )
f x1 , x 2 , x3 = c 3 e
(
c x1 + x2 + x 3 )
( )
I A x1 , x 2 , x3 ,
where
( ) (
A = 0, 0, 0, , ) ( )
i) Determine the constant c;
ii) Find the marginal p.d.f. of each one of the r.v.s Xj, j = 1, 2, 3;
iii) Find the conditional (joint) p.d.f. of X1, X2, given X3, and the conditional
p.d.f. of X1, given X2, X3;
iv) Find the conditional d.f.s corresponding to the conditional p.d.f.s in (iii).
4.2.6 Consider the function given below:
yx e y
( )
f x y = x!
0,
, x = 0, 1, ; y 0

otherwise.
4.3
4.1 Quantiles
The Cumulative
and Modes
Distribution
of a Distribution
Function 99

i) Show that for each fixed y, f(|y) is a p.d.f., the conditional p.d.f. of an r.v.
X, given that another r.v. Y equals y;
ii) If the marginal p.d.f. of Y is Negative Exponential with parameter = 1,
what is the joint p.d.f. of X, Y?
iii) Show that the marginal p.d.f. of X is given by f(x) = ( 12 )x+1 IA(x), where
A = {0, 1, 2, . . . }.
4.2.7 Let Y be an r.v. distributed as P() and suppose that the conditional
distribution of the r.v. X, given Y = n, is B(n, p). Determine the p.d.f. of X and
the conditional p.d.f. of Y, given X = x.
4.2.8 Consider the function f defined as follows:
x 12 + x 22
( )
f x1 , x 2 =
1
2
exp
2 4e
+
1 3 3
(
x1 x 2 I [ 1, 1][ 1, 1] x1 , x 2 )

and show that:
i) f is a non-Normal Bivariate p.d.f.
ii) Both marginal p.d.f.s

( ) ( )

f1 x1 = f x1 , x 2 dx 2

and

( ) ( )

f2 x 2 = f x1 , x 2 dx1

are Normal p.d.f.s.

4.3 Quantiles and Modes of a Distribution


Let X be an r.v. with d.f. F and consider a number p such that 0 < p < 1. A
pth quantile of the r.v. X, or of its d.f. F, is a number denoted by xp and having
the following property: P(X xp) p and P(X xp) 1 p. For p = 0.25 we
get a quartile of X, or its d.f., and for p = 0.5 we get a median of X, or its
d.f. For illustrative purposes, consider the following simple examples.
EXAMPLE 3 Let X be an r.v. distributed as U(0, 1) and let p = 0.10, 0.20, 0.30, 0.40, 0.50,
0.60, 0.70, 0.80 and 0.90. Determine the respective x0.10, x0.20, x0.30, x0.40, x0.50, x0.60,
x0.70, x0.80, and x0.90.
Since for 0 x 1, F(x) = x, we get: x0.10 = 0.10, x0.20 = 0.20, x0.30 = 0.30,
x0.40 = 0.40, x0.50 = 0.50, x0.60 = 0.60, x0.70 = 0.70, x0.80 = 0.80, and x0.90 = 0.90.
EXAMPLE 4 Let X be an r.v. distributed as N(0, 1) and let p = 0.10, 0.20, 0.30, 0.40, 0.50,
0.60, 0.70, 0.80 and 0.90. Determine the respective x0.10, x0.20, x0.30, x0.40, x0.50, x0.60,
x0.70, x0.80, and x0.90.
100 4 Distribution Functions, Probability Densities, and Their Relationship

Typical cases:
F(x) F(x)
1
p

p
x [ ] x
xp 0 0
xp
(a) (b)
F(x)

x
0 xp
(c)
F(x) F(x)

1
1
p
p

x [ ] x
0 xp 0
xp
(d) (e)
Figure 4.3 Observe that the figures demonstrate that, as defined, xp need not be unique.

From the Normal Tables (Table 3 in Appendix III), by linear interpolation


and symmetry, we find: x0.10 = 1.282, x0.20 = 0.842, x0.30 = 0.524, x0.40 = 0.253,
x0.50 = 0, x0.60 = 0.253, x0.70 = 0.524, x0.80 = 0.842, and x0.90 = 1.282.
Knowledge of quantiles xp for several values of p provides an indication as
to how the unit probability mass is distributed over the real line. In Fig. 4.3
various cases are demonstrated for determining graphically the pth quantile of
a d.f.
Let X be an r.v. with a p.d.f. f. Then a mode of f, if it exists, is any number
which maximizes f(x). In case f is a p.d.f. which is twice differentiable, a mode
can be found by differentiation. This process breaks down in the discrete cases.
The following theorems answer the question for two important discrete cases.
4.3
4.1 Quantiles
The Cumulative
and Modes
Distribution
of a Distribution
Function 101

THEOREM 5 Let X be B(n, p); that is,


n
()
f x = px qn x ,
x
0 < p < 1, q = 1 p, x = 0, 1, , n.

Consider the number (n + 1)p and set m = [(n + 1)p], where [y] denotes the
largest integer which is y. Then if (n + 1)p is not an integer, f(x) has a unique
mode at x = m. If (n + 1)p is an integer, then f(x) has two modes obtained for
x = m and x = m 1.
PROOF For x 1, we have

n x n x
p q
f x
=
() x
(
f x1

)
n x 1 n x +1
p q
x 1
n!
p x q n x
=
(
x! n x ! ) =
nx +1 p
.
n! x q
p x 1 q n x +1
( )(
x 1! nx +1! )
That is,
f x() =
nx +1 p
.
(
f x1 ) x q

Hence f(x) > f(x 1) ( f is increasing) if and only if

(n x + 1) p > x(1 p), or np xp + p > x xp, or (n + 1) p > x.


Thus if (n + 1)p is not an integer, f(x) keeps increasing for x m and then
decreases so the maximum occurs at x = m. If (n + 1)p is an integer, then the
maximum occurs at x = (n + 1)p, where f(x) = f(x 1) (from above calcula-
tions). Thus

(
x = n+1 p1 )
is a second point which gives the maximum value. "
THEOREM 6 Let X be P(); that is,
x
()
f x = e
x!
, x = 0, 1, 2, , > 0.

Then if is not an integer, f(x) has a unique mode at x = []. If is an integer,


then f(x) has two modes obtained for x = and x = 1.
PROOF For x 1, we have
102 4 Distribution Functions, Probability Densities, and Their Relationship

f x () =
( )
e x x!
=

.
(
f x1 ) e
[ (x 1)!]
x 1 x

Hence f(x) > f(x 1) if and only if > x. Thus if is not an integer, f(x) keeps
increasing for x [] and then decreases. Then the maximum of f(x) occurs
at x = []. If is an integer, then the maximum occurs at x = . But in this case
f(x) = f(x 1) which implies that x = 1 is a second point which gives
the maximum value to the p.d.f. "

Exercises
4.3.1 Determine the pth quantile xp for each one of the p.d.f.s given in
Exercises 3.2.1315, 3.3.1316 (Exercise 3.2.14 for = 14 ) in Chapter 3 if p =
0.75, 0.50.
4.3.2 Let X be an r.v. with p.d.f. f symmetric about a constant c (that is,
f(c x) = f(c + x) for all x ! ). Then show that c is a median of f.
4.3.3 Draw four graphstwo each for B(n, p) and P()which represent
the possible occurrences for modes of the distributions B(n, p) and P().
4.3.4 Consider the same p.d.f.s mentioned in Exercise 4.3.1 from the point
of view of a mode.

4.4* Justification of Statements 1 and 2


In this section, a rigorous justification of Statements 1 and 2 made in Section
4.1 will be presented. For this purpose, some preliminary concepts and results
are needed and will be also discussed.
DEFINITION 1 A set G in ! is called open if for every x in G there exists an open interval
containing x and contained in G. Without loss of generality, such intervals may
be taken to be centered at x.
It follows from this definition that an open interval is an open set, the
entire real line ! is an open set, and so is the empty set (in a vacuous manner).
LEMMA 1 Every open set in ! is measurable.
PROOF Let G be an open set in !, and for each x G, consider an open
interval centered at x and contained in G. Clearly, the union over x, as x varies
in G, of such intervals is equal to G. The same is true if we consider only those
intervals corresponding to all rationals x in G. These intervals are countably
many and each one of them is measurable; then so is their union. "
4.1
4.4*The
Justification
CumulativeofDistribution
StatementsFunction
1 and 2 103

DEFINITION 2 A set G in ! m, m 1, is called open if for every x in G there exists an open cube
in ! m containing x and contained in G; by the term open cube we mean the
Cartesian product of m open intervals of equal length. Without loss of gener-
ality, such cubes may be taken to be centered at x.
LEMMA 2 Every open set in ! n is measurable.
PROOF It is analogous to that of Lemma 1. Indeed, let G be an open set in
! m, and for each x G, consider an open cube centered at x and contained in
G. The union over x, as x varies in G, of such cubes clearly is equal to G. The
same is true if we restrict ourselves to xs in G whose m coordinates are
rationals. Then the resulting cubes are countably many, and therefore their
union is measurable, since so is each cube. "
DEFINITION 3 Recall that a function g: S ! ! is said to be continuous at x0 S if for
every > 0 there exists a = (, x0) > 0 such that |x x0| < implies |g(x) g(x0)|
< . The function g is continuous in S if it is continuous for every x S.
It follows from the concept of continuity that 0 implies 0.
LEMMA 3 Let g: ! ! be continuous. Then g is measurable.
PROOF By Theorem 5 in Chapter 1 it suffices to show that g1(G) are meas-
urable sets for all open intevals G in !. Set B = g1(G). Thus if B = , the
assertion is valid, so let B and let x0 be an arbitrary point of B, so that g(x0)
G. Continuity of g at x0 implies that for every > 0 there exists = (, x0)
> 0 such that |x x0| < implies |g(x) g(x0)| < . Equivalently, x (x0 , x0
+ ) implies g(x) (g(x0) , g(x0) + ). Since g(x0) G and G is open, by
choosing sufficiently small, we can make so small that (g(x0) , g(x0) + )
is contained in G. Thus, for such a choice of and , x (x0 , x0 + ) implies
that (g(x0) , g(x0) + ) G. But B(= g1(G)) is the set of all x ! for which
g(x) G. As all x (x0 , x0 + ) have this property, it follows that (x0 , x0
+ ) B. Since x0 is arbitrary in B, it follows that B is open. Then by Lemma
1, it is measurable. "
The concept of continuity generalizes, of course, to Euclidean spaces of
higher dimensions, and then a result analogous to the one in Lemma 3 also
holds true.

DEFINITION 4 A function g : S ! k ! m (k, m 1) is said to be continuous at x0 ! k if for


every > 0 there exists a = (, x0) > 0 such that ||x x0|| < implies ||g(x)
g(x0)|| < . The function g is continuous in S if it is continuous for every x S.
Here ||x|| stands for the usual norm in ! k; i.e., for x = (x1, . . . , xk), ||x|| =
(ik=1 xi2 ) , and similarly for the other quantities.
12

Once again, from the concept of continuity it follows that 0 implies


0.
LEMMA 4 Let g: ! k ! m be continuous. Then g is measurable.
104 4 Distribution Functions, Probability Densities, and Their Relationship

PROOF The proof is similar to that of Lemma 3. The details are presented
here for the sake of completeness. Once again, it suffices to show that g1(G)
are measurable sets for all open cubes G in !m. Set B = g1(G). If B = the
assertion is true, and therefore suppose that B and let x0 be an arbitrary
point of B. Continuity of g at x0 implies that for every > 0 there exists a =
(, x0) > 0 such that ||x x0|| < implies ||g(x) g(x0)|| < ; equivalently, x
S(x0, ) implies g(x) S(g(x0), ), where S(c, r) stands for the open sphere with
center c and radius r. Since g(x0) G and G is open, we can choose so small
that the corresponding is sufficiently small to imply that g(x) S(g(x0), ).
Thus, for such a choice of and , x S(x0, ) implies that g(x) S(g(x0), ).
Since B(= g1(G)) is the set of all x ! k for which g(x) G, and x S(x0, )
implies that g(x) S(g(x0), ), it follows that S(x0, ) B. At this point, observe
that it is clear that there is a cube containing x0 and contained in S(x0, ); call
it C(x0, ). Then C(x0, ) B, and therefore B is open. By Lemma 2, it is also
measurable. "
We may now proceed with the justification of Statement 1.
THEOREM 7 Let X : (S, A) (! k, B k) be a random vector, and let g : (! k, B k) (! m, Bm)
be measurable. Then g(X): (S, A) (! m, Bm) and is a random vector. (That
is, measurable functions of random vectors are random vectors.)
PROOF To prove that [g(X)]1(B) A if B B m, we have

[g(X)] (B) = X [g (B)] = X (B ), ( )


1
1 1 1
1 where B1 = g 1 B B k
1
by the measurability of g. Also, X (B1) A since X is measurable. The proof
is completed. "
To this theorem, we have the following
COROLLARY Let X be as above and g be continuous. Then g(X) is a random vector. (That
is, continuous functions of random vectors are random vectors.)
PROOF The continuity of g implies its measurability by Lemma 3, and there-
fore the theorem applies and gives the result. "
DEFINITION 5 For j = 1, . . . , k, the jth projection function gj is defined by: gj: ! k ! and
gj(x) = gj(x1, . . . , xk) = xj.
It so happens that projection functions are continuous; that is,
LEMMA 5 The coordinate functions gj, j = 1, . . . , k, as defined above, are continuous.
PROOF For an arbitrary point x0 in ! K, consider x ! K such that ||x x0|| <
for some > 0. This is equivalent to ||x x0||2 < 2 or j =1 ( x j x ) < 2 which
k 2
0j

implies that (xj x0j)2 < 2 for j = 1, . . . , k, or |xj x0j| < , j = 1, . . . , k. This last
expression is equivalent to |gj(x) gj(x0)| < , j = 1, . . . , k. Thus the definition
of continuity of gj is satisfied here for = . "
Now consider a k-dimensional function X defined on the sample space S.
Then X may be written as X = (X1, . . . , Xk), where Xj, j = 1, . . . , k are real-
valued functions. The question then arises as to how X and Xj, j = 1, . . . , k are
4.1 The Cumulative DistributionExercises
Function 105

related from a measurability point of view. To this effect, we have the follow-
ing result.
Let X = (X1, . . . , Xk) : (S, A) (! k, B k). Then X is an r. vector if and only if
THEOREM 8 Xj, j = 1, . . . , k are r.v.s.
PROOF Suppose X is an r. vector and let gj, j = 1, . . . , k be the coordinate
functions defined on ! k. Then gjs are continuous by Lemma 5 and therefore
measurable by Lemma 4. Then for each j = 1, . . . , k, gj(X) = gj(X1, . . . , Xk) =
Xj is measurable and hence an r.v.
Next, assume that Xj, j = 1, . . . , k are r.v.s. To show that X is an r. vector,
by special case 3 in Section 2 of Chapter 1, it suffices to show that X1(B) A
for each B = (, x1] (, xk], x1, . . . , xk !. Indeed,
k

X 1 ( B) = (X B) = ( X j ( , x j ], j = 1, , k ) = I X j1 (( , x j ]) A .
j =1

The proof is completed. "

Exercises
4.4.1 If X and Y are functions defined on the sample space S into the real line
!, show that:

{s S ; X (s) + Y (s) < x} = U [{s S ; X (s) < r} {s S ; Y (s) < x r}],


r Q

where Q is the set of rationals in !.


4.4.2 Use Exercise 4.4.1 in order to show that, if X and Y are r.v.s, then so
is the function X + Y.
4.4.3
ii) If X is an r.v., then show that so is the function X.
ii) Use part (i) and Exercise 4.4.2 to show that, if X and Y are r.v.s, then so is
the function X Y.
4.4.4
ii) If X is an r.v., then show that so is the function X 2.
1
ii) Use the identity: XY = 2 (X + Y)2 12 (X 2 + Y 2) in conjunction with part (i)
and Exercises 4.4.2 and 4.4.3(ii) to show that, if X and Y are r.v.s, then so
is the function XY.
4.4.5
ii) If X is an r.v., then show that so is the function 1
X
, provided X 0.
ii) Use part (i) in conjunction with Exercise 4.4.4(ii) to show that, if X and Y
are r.v.s, then so is the function XY , provided Y 0.
106 5 Moments of Random VariablesSome Moment and Probability Inequalities

Chapter 5

Moments of Random VariablesSome


Moment and Probability Inequalities

5.1 Moments of Random Variables


In the definitions to be given shortly, the following remark will prove useful.
REMARK 1 We say that the (infinite) series xh(x), where x = (x1, . . . , xk)
varies over a discrete set in !k, k 1, converges absolutely if x|h(x)| < . Also

( )
we say that the integral h x1 , , xk dx1 dxk converges absolutely
if

h( x1 , x2 , , xk ) dx1 dx2 dxk < .


In what follows, when we write (infinite) series or integrals it will always be


assumed that they converge absolutely. In this case, we say that the moments to
be defined below exist.
Let X = (X1, . . . , Xk) be an r. vector with p.d.f. f and consider the (measurable)
function g: ! k !, so that g(X) = g(X1, . . . , Xk) is an r.v. Then we give the
DEFINITION 1 iii) For n = 1, 2, . . . , the nth moment of g(X) is denoted by E[g(X)]n and is
defined by:

[ ( )] ( ) (

)
n
g x f x , x = x1 , , xk
[ ( )]
n
E g X = x
[(
g x ,
)] ( )
n
1 , xk f x1 , , xk dx1 dxk .

For n = 1, we get
g x f x ()()
[ ( )]

E g X = x
( )( )

g x1 , , xk f x1 , , xk dx1 dxk

106
5.1 Moments of Random Variables 107

and call it the mathematical expectation or mean value or just mean of g(X).
Another notation for E[g(X)] which is often used is g(X), or [g(X)], or just
, if no confusion is possible.
iii) For r > 0, the rth absolute moment of g(X) is denoted by E|g(X)|r and is
defined by:

() () ( )
r
g x f x , x = x1 , , xk
( )
r
Eg X = x
g x ,
( ) ( )
r
1 , xk f x1 , , xk dx1 dxk .

iii) For an arbitrary constant c, and n and r as above, the nth moment and rth
absolute moment of g(X) about c are denoted by E[g(X) c]n, E|g(X) c|r,
respectively, and are defined as follows:

[() ] () (
)
n
g x c f x , x = x1 , , xk
[( ) ]
n
E g X c = x
[( ) ] (
g x ,
)
n
1 , xk c f x1 , , xk dx1 dxk ,

and

() () ( )
r
g x c f x , x = x1 , , xk
( )
r
Eg X c = x
g x ,
( ) ( )
r
1 , xk c f x1 , , xk dx1 dxk .

For c = E[g(X)], the moments are called central moments. The 2nd central
moment of g(X), that is,

{( ) [ ( )]}
2
E g X E g X

[ g( x) Eg(X )] f ( x),

( )
2
x = x1 , , xk
= x
g x ,
[( ) ( )] f ( x , , x )dx
2
1 , xk Eg X 1 k 1 dxk

is called the variance of g(X) and is also denoted by 2[g(X)], or g2( X ), or just
2, if no confusion is possible. The quantity + 2 [ g(X )] = [ g(X )] is called the
standard deviation (s.d.) of g(X) and is also denoted by g(X), or just , if no
confusion is possible. The variance of an r.v. is referred to as the moment of
inertia in Mechanics.

5.1.1 Important Special Cases


1. Let g(X1, . . . , Xk) = X 1 X k , where nj 0 are integers. Then
n 1 n k

E( X 1n X kn ) is called the (n1, . . . , nk)-joint moment of X1, . . . , Xk. In par-


1 k

ticular, for n1 = = nj1 = nj+1 = = nk = 0, nj = n, we get


108 5 Moments of Random VariablesSome Moment and Probability Inequalities

()
x nj f x =
x
(
x nj f x1 , , xk )
( )

E X n
= ( x , ,x )
1 k
j
xn f x ,
( )

j 1 , xk dx1 dxk

( )
x nj f j x j
x
= j

j j j ( )
x n f x dx
j

which is the nth moment of the r.v. Xj. Thus the nth moment of an r.v. X with
p.d.f. f is

x n f x ()
( )
E Xn

= x
()

x n f x dx.

For n = 1, we get

xf x

()
( )
E X = x
()

xf x dx

which is the mathematical expectation or mean value or just mean of X. This


quantity is also denoted by X or (X) or just when no confusion is possible.
The quantity X can be interpreted as follows: It follows from the defini-
tion that if X is a discrete uniform r.v., then X is just the arithmetic average of
the possible outcomes of X. Also, if one recalls from physics or elementary
calculus the definition of center of gravity and its physical interpretation as the
point of balance of the distributed mass, the interpretation of X as the mean
or expected value of the random variable is the natural one, provided the
probability distribution of X is interpreted as the unit mass distribution.
REMARK 2 In Definition 1, suppose X is a continuous r.v. Then E[g(X)] =
g( x)f ( x)dx. On the other hand, from the last expression above, E(X) =

xf ( x)dx. There seems to be a discrepancy between these two definitions.


More specifically, in the definition of E[g(X)], one would expect to use the
p.d.f. of g(X) rather than that of X. Actually, the definition of E[g(X)], as
given, is correct and its justification is roughly as follows: Consider E[g(x)] =
g( x)f ( x)dx and set y = g(x). Suppose that g is differentiable and has an

inverse g1, and that some further conditions are met. Then

g( x) f ( x)dx = yf [g ( y)] dy g ( y) dy.



1 1 d
5.1 Moments of Random Variables 109

On the other hand, if fY is the p.d.f. of Y, then fY ( y) = f [ g 1 ( y)] dyd g 1 ( y) .


Therefore the last integral above is equal to yfY ( y)dy, which is consonant

with the definition of E ( X ) = xf ( x) dx. (A justification of the above deriva-


tions is given in Theorem 2 of Chapter 9.)


2. For g as above, that is, g(X1, . . . , Xk) = X 1n X kn and n1 = = nj1 = 1 k

nj+1 = = nk = 0, nj = 1, and c = E(Xj), we get



( ) () ( )
n
x j EX j f x , x = x1 , , xk
( )
n
E X j EX j = x
(
x EX n f x ,
j j ) (
1 , xk dx1 dxk )

( ) ( )
n

x j EX j f j x j
= xj
(
x EX n f x dx
j j j j j ) ( )
which is the nth central moment of the r.v. Xj (or the nth moment of Xj about its
mean).
Thus the nth central moment of an r.v. X with p.d.f. f and mean is

( ) f (x) = (x ) f (x)
n n

x EX
( ) ( )
n n
E X EX =E X = x x

x EX ( ) f (x)dx = (x ) f (x)dx.
n n

In particular, for n = 2 the 2nd central moment of X is denoted by X2 or 2(X)


or just 2 when no confusion is possible, and is called the variance of X. Its
positive square root X or (X) or just is called the standard deviation (s.d.)
of X.
As in the case of X, X2 has a physical interpretation also. Its definition
corresponds to that of the second moment, or moment of inertia. One recalls
that a large moment of inertia means the mass of the body is spread widely
about its center of gravity. Likewise a large variance corresponds to a probabil-
ity distribution which is not well concentrated about its mean value.
3. For g(X1, . . . , Xk) = (X1 EX1)n (Xk EXk)n , the quantity 1 k

(
E X1 EX1 X k EX k ) ( )
n n 1 k


is the (n1, . . . , nk)-central joint moment of X1, . . . , Xk or the (n1, . . . , nk)-joint
moment of X1, . . . , Xk about their means.
4. For g(X1, . . . , Xk) = Xj(Xj 1) (Xj n + 1), j = 1, . . . , k, the quantity

x j x j 1 x j n + 1 f j x j ( ) ( ) ( )
[ ( )
x
E Xj Xj 1 Xj n+1 = ( )] j

x x 1 x n + 1 f x dx
j j j j j j ( ) ( ) ( )
110 5 Moments of Random VariablesSome Moment and Probability Inequalities

is the nth factorial moment of the r.v. Xj. Thus the nth factorial moment of an
r.v. X with p.d.f. f is

(
x x 1 x n + 1 f x ) ( )()
[ ( )
(
E X X 1 X n+1 = x )]
( ) ( )()

x x 1 x n + 1 f x dx.

5.1.2 Basic Properties of the Expectation of an R.V.
From the very definition of E[g(X)], the following properties are immediate.
(E1) E(c) = c, where c is a constant.
(E2) E[cg(X)] = cE[g(X)], and, in particular, E(cX) = cE(X) if X is an
r.v.
(E3) E[g(X) + d] = E[g(X)] + d, where d is a constant. In particular,
E(X + d) = E(X) + d if X is an r.v.
(E4) Combining (E2) and (E3), we get E[cg(X) + d] = cE[g(X)] + d,
and, in particular, E(cX + d) = cE(X) + d if X is an r.v.
[ ( )] [ ( )]
(E4) E nj = 1 cj g j X = nj = 1 cj E g j X .
In fact, for example, in the continuous case, we have

n n
( ) ( ) ( )

E c j g j X = c j g j x1 , , xk f x1 , , xk dx1 dxk
j =1
j =1
n

( )( )

= c j g j x1 , , xk f x1 , , xk dx1 dxk

j =1

[ ( )]
n
= cjE gj X .
j =1

The discrete case follows similarly. In particular,


(E4) E ( jn= 1 c j X j ) = jn= 1 c j E ( X j ).
(E5) If X 0, then E(X) 0.
Consequently, by means of (E5) and (E4), we get that
(E5) If X Y, then E(X) E(Y), where X and Y are r.v.s (with finite
expectations).
(E6) |E[g(X)]| E|g(X)|.
(E7) If E|X|r < for some r > 0, where X is an r.v., then E|X|r< for
all 0 < r < r.
This is a consequence of the obvious inequality |X|r 1 + |X|r and (E5).
Furthermore, since of n = 1, 2, . . . , we have |Xn| = |X|n, by means of (E6),
it follows that
5.1 Moments of Random Exercises
Variables 111

(E7) If E(Xn) exists (that is, E|X|n < ) for some n = 2, 3, . . . , then E(Xn)
also exists for all n = 1, 2, . . . with n < n.

5.1.3 Basic Properties of the Variance of an R.V.


Regarding the variance, the following properties are easily established by
means of the definition of the variance.
(V1) 2(c) = 0, where c is a constant.
(V2) 2[cg(X)] = c2 2[g(X)], and, in particular, 2(cX) = c2 2(X), if X is
an r.v.
(V3) 2[g(X) + d] = 2[g(X)], where d is a constant. In particular,
2(X + d) = 2(X), if X is an r.v.
In fact,

[( ) ] {[ ( ) ] [ ( ) ]}
2
2 g X +d = E g X +d E g X +d

= E[ g( X) Eg( X)] = [ g( X)].


2
2

(V4) Combining (V2) and (V3), we get 2[cg(X) + d] = c2 2[g(X)],


and, in particular, 2(cX + d) = c2 2(X), if X is an r.v.
(V5) 2[g(X)] = E[g(X)]2 [Eg(X)]2, and, in particular,
(V5) 2(X) = E(X 2) (EX)2, if X is an r.v.
In fact,

[ ( )] [ ( ) ( )] = E[g(X)] 2 g(X)Eg(X) + [Eg(X)]


2 2 2
2 g X = E g X Eg X

= E[ g( X)] 2[ Eg( X)] + [ Eg( X)] = E[ g( X)] [ Eg( X)] ,


2 2 2 2 2

the equality before the last one being true because of (E4).
(V6) 2(X) = E[X(X 1)] + EX (EX)2, if X is an r.v., as is easily seen.
This formula is especially useful in calculating the variance of a
discrete r.v., as is seen below.

Exercises
5.1.1 Verify the details of properties (E1)(E7).
5.1.2 Verify the details of properties (V1)(V5).
5.1.3 For r < r, show that |X|r 1 + |X|r and conclude that if E|X|r < , then
E|X|r for all 0 < r < r.
112 5 Moments of Random VariablesSome Moment and Probability Inequalities


5.1.4 Verify the equality (E[ g( X )] =) g( x) fX ( x)dx = yfY ( y)dy for the
case that X N(0, 1) and Y = g(X) = X2.
5.1.5 For any event A, consider the r.v. X = IA, the indicator of A defined by
IA(s) = 1 for s A and IA(s) = 0 for s Ac, and calculate EXr, r > 0, and also
2(X).
5.1.6 Let X be an r.v. such that

( ) (
P X = c = P X = c = .
2
1
)
Calculate EX, 2(X) and show that

( ).
2 X
(
P X EX c = ) c2
5.1.7 Let X be an r.v. with finite EX.
ii) For any constant c, show that E(X c)2 = E(X EX)2 + (EX c)2;
ii) Use part (i) to conclude that E(X c)2 is minimum for c = EX.
5.1.8 Let X be an r.v. such that EX4 < . Then show that
ii) E(X EX)3 = EX3 3(EX)(EX)2 + 2(EX)3;
ii) E(X EX)4 = EX4 4(EX)(EX3) + 6(EX)2(EX2) 3(EX)4.
5.1.9 If EX4 < , show that:

[ ( )] [ ( )( )]
E X X 1 = EX 2 EX ; E X X 1 X 2 = EX 3 3EX 2 + 2EX ;

E[ X ( X 1)( X 2)( X 3)] = EX 6EX + 11EX


4 3 2
6EX .

(These relations provide a way of calculating EXk, k = 2, 3, 4 by means of


the factorial moments E[X(X 1)], E[X(X 1)(X 2)], E[X(X 1)(X 2)
(X 3)].)
5.1.10 Let X be the r.v. denoting the number of claims filed by a policy-
holder of an insurance company over a specified period of time. On the basis
of an extensive study of the claim records, it may be assumed that the distribu-
tion of X is as follows:

x 0 1 2 3 4 5 6

f(x) 0.304 0.287 0.208 0.115 0.061 0.019 0.006

iii) Calculate the EX and the 2(X);


iii) What premium should the company charge in order to break even?
iii) What should be the premium charged if the company is to expect to come
ahead by $M for administrative expenses and profit?
5.1 Moments of Random Exercises
Variables 113

5.1.11 A roulette wheel has 38 slots of which 18 are red, 18 black, and 2
green.
iii) Suppose a gambler is placing a bet of $M on red. What is the gamblers
expected gain or loss and what is the standard deviation?
iii) If the same bet of $M is placed on green and if $kM is the amount
the gambler wins, calculate the expected gain or loss and the standard
deviation.
iii) For what value of k do the two expectations in parts (i) and (ii) coincide?
iv) Does this value of k depend on M?
iv) How do the respective standard deviations compare?
5.1.12 Let X be an r.v. such that P(X = j) = ( 12 )j, j = 1, 2, . . . .
ii) Compute EX, E[X(X 1)];
ii) Use (i) in order to compute 2(X).
5.1.13 If X is an r.v. distributed as U(, ), show that

( )
2
+
EX =
2
, X2
( ) =
12
.

5.1.14 Let the r.v. X be distributed as U(, ). Calculate EXn for any positive
integer n.
5.1.15 Let X be an r.v. with p.d.f. f symmetric about a constant c (that is,
f(c x) = f(c + x) for every x).
ii) Then if EX exists, show that EX = c;
ii) If c = 0 and EX2n+1 exists, show that EX2n+1 = 0 (that is, those moments of X
of odd order which exist are all equal to zero).
5.1.16 Refer to Exercise 3.3.13(iv) in Chapter 3 and find the EX for those s
for which this expectation exists, where X is an r.v. having the distribution in
question.
5.1.17 Let X be an r.v. with p.d.f. given by

x
()
f x = ()
I c ,c x .
c2 ( )

Compute EXn for any positive integer n, E|Xr|, r > 0, 2(X).


5.1.18 Let X be an r.v. with finite expectation and d.f. F.
ii) Show that

[ ( )] ()
0
EX = 1 F x dx F x dx;
0
114 5 Moments of Random VariablesSome Moment and Probability Inequalities

ii) Use the interpretation of the definite integral as an area in order to give a
geometric interpretation of EX.
5.1.19 Let X be an r.v. of the continuous type with finite EX and p.d.f. f.
ii) If m is a median of f and c is any constant, show that

( )()
c
E X c = E X m + 2 c x f x dx;
m

ii) Utilize (i) in order to conclude that E|X c| is minimized for c = m. (Hint:
Consider the two cases that c m and c < m, and in each one split the
integral from to c and c to in order to remove the absolute value.
Then the fact that f ( x)dx = m f ( x)dx = 21 and simple manipulations
m

prove part (i). For part (ii), observe that m (c x) f ( x)dx 0 whether c m
c

or c < m.)
5.1.20 If the r.v. X is distributed according to the Weibull distribution (see
Exercise 4.1.15 in Chapter 4), then:
1 1 2 2
ii) Show that EX = 1 + , EX 2 = 1 + , so that

2 1
( )
2 X = 1 + 2 1 + 2 ,


where recall that the Gamma function is defined by = 0 t 1 e t dt , ()
> 0;
ii) Determine the numerical values of EX and 2(X) for = 1 and = 12 ,
= 1 and = 2.

5.2 Expectations and Variances of Some r.v.s


5.2.1 Discrete Case
1. Let X be B(n, p). Then E(X) = np, 2(X) = npq. In fact,

n
n n

( )
E X = x p x q n x = x
n!
px qn x
x = 0 x x =1 x! n x ! ( )
=
n
( )
n n1 !
px qn x
( x 1)!(n x)!
x =1

= np
n
(n 1)! p q
( )( ) x 1 n 1 x 1

( ) [( ) ( )]
x
x =11 ! n 1 x 1 !

(n 1)! p q( ) = np p + q
n 1

( )
n 1
= np
n 1 x
x
= np.
x![( n 1) x]!
x= 0
5.2 Expectations
5.1 Moments
and Variances
of Random
of Some
Variables
R.V.s 115

Next,
[ (
E X X 1 )]
n
= x x1 ( ) x! nn! x ! p q x n x

x= 0 ( )
n n( n 1)( n 2)!
= x( x 1)
( 2 x2 )(
n 2 x2 )
p p q
x=2 x( x 1)( x 2)![( n 2) ( x 2)]!

= n( n 1) p 2
n
(n 2)! p q
( )( ) x2 n 2 x2

( x 2)![(n 2) ( x 2)]!
x=2

= n( n 1) p 2 (n 2)! p q( )
n 2
x n 2 x

x![( n 2) x]!
x= 0

= n( n 1) p ( p + q) = n( n 1) p .
2 n 2 2

That is,

[ (
E X X 1 = n n 1 p2 . )] ( )
Hence, by (V6),

( ) [ ( )] ( ) ( )
2
2 X = E X X 1 + EX EX = n n 1 p 2 + np n 2 p 2
2 2 2
= n p np + np n p = np 1 p = npq. 2 2
( )
2. Let X be P(). Then E(X) = 2(X) = . In fact,


x
x
x 1
( )
E X = xe = xe = e
x= 0 x! x =1 x x1 ! (
x =1 x 1 ! ) ( )

x
= e = e e = .
x = 0 x!

Next,

[ ( )] x

E X X 1 = x x 1 e
x= 0
( ) x!

x
x
= x x 1 e ( ) = 2 e = 2 .
x=2 (
x x1 x 2 ! )( x = 0 x! )
Hence EX2 = 2 + , so that, 2(X) = 2 + 2 = .
REMARK 3 One can also prove that the nth factorial moment of X is n; that
is, E[X(X 1) (X n + 1)] = n.
116 5 Moments of Random VariablesSome Moment and Probability Inequalities

5.2.2 Continuous Case


1. Let X be N(0, 1). Then

( ) 2( (n)!) ,
2n !
( )
E X 2 n+1 = 0, E X 2 n = n
n 0.

In particular, then

( )
E X = 0, ( )
2 X = E X2 = ( ) 2
2 1!
= 1.

In fact,

( ) 1

2
E X 2 n+1 = x 2 n+1 e x 2
dx.

2
But
0
x dx = x 2n +1e x 2 dx + x 2n +1e x 2 dx
2 n +1 x 2 2 2 2
e
0
0
= y 2n +1e y 2 dy + x 2n +1e x 2 dx
2 2

0

= x 2n +1e x 2 dx + x 2n +1e x 2 dx = 0.
2 2

0 0

Thus E(X2n+1) = 0. Next,



x dx = 2 x 2 n e x
2 2
2n
ex 2 2
dx,
0

as is easily seen, and



x 2n e x 2 dx = x 2n 1de x
2 2
2
0 0

( )
2 2
= x 2n 1e x 2
+ 2n 1 x 2n 2 e x 2 dx
0 0

( )
2
= 2n 1 x 2n 2 e x 2 dx,
0

and if we set m2n = E(X2n), we get then

( )
m2 n = 2 n 1 m2 n 2 , and similarly,
m 2 n 2 = (2 n 3)m 2 n 4

M
m2 = 1 m0

(
m0 = 1 since m0 = E X 0 = E 1 = 1 . ( ) () )
Multiplying them out, we obtain
5.2 Expectations
5.1 Moments
and Variances
of Random
of Some
Variables
R.V.s 117

( )( )
m2 n = 2 n 1 2 n 3 1
1 2 ( 2 n 3)( 2 n 2)( 2 n 1)( 2 n) (2 n)!
= =
2 ( 2 n 2)( 2 n) (2 1) [2( n 1)](2 n)
=
(2 n)! =
(2 n)! .
2 [1 ( n 1) n] 2 ( n!)
n n

REMARK 4 Let now X be N(, 2). Then (X )/ is N(0, 1). Hence

X 2 X
E = 0, = 1.

But

X 1
E

= EX .

( )
Hence


1

( )
E X = 0,

so that E(X) = . Next,

X
2

1 2
= 2 X ( )
and then

1 2
2
( )
X = 1,

so that 2(X) = 2.
2. Let X be Gamma with parameters and . Then E(X) = and
2(X) = 2. In fact,

( ) 1 1

( ) ( )
EX = xx 1 e x dx = x e x dx
0
0

x e x x 1 e x dx
( ) 0
= x de x =
0
( )
0

( )
= x 1 e x dx = .
0
118 5 Moments of Random VariablesSome Moment and Probability Inequalities

Next,

( ) (1) ( )

E X2 = 0 x +1 e x dx = 2 + 1

and hence

( ) ( )
2 X = 2 + 1 2 2 = 2 + 1 = 2 . ( )
REMARK 5
ii) If X is r2 , that is, if = r/2, = 2, we get E(X) = r, 2(X) = 2r.
ii) If X is Negative Exponential, that is, if = 1, = 1/, we get E(X) = 1/,
2(X) = 1/2.
3. Let X be Cauchy. Then E(Xn) does not exist for any n 1. For example,
for n = 1, we get

x dx
I= .

( )
2
2 + x

For simplicity, we set = 0, = 1 and we have

1 x dx

1 1 d x
2
( )
1 + x 2 2 1 + x 2
I= =

1 1 d 1+ x
2
( 1 ) ( )

=
2 1+ x 2
=
2
log 1 + x 2

=
1
2
(
, )
which is an indeterminate form. Thus the Cauchy distribution is an example of
a distribution without a mean.
REMARK 6 In somewhat advanced mathematics courses, one encounters
sometimes the so-called Cauchy Principal Value Integral. This coincides with
the improper Riemann integral when the latter exists, and it often exists even
if the Riemann integral does not. It is an improper integral in which the limits
are taken symmetrically. As an example, for = 1, = 0, we have, in terms of
the principal value integral,

I * = lim
A
1 A xdx
A 1+ x 2
=
1
lim log 1 + x 2
2 A
( ) A
A

=
1
[ ( )
lim log 1 + A 2 log 1 + A 2 = 0.
2 A
( )]
5.1 Moments of Random Exercises
Variables 119

Thus the mean of the Cauchy exists in terms of principal value, but not in the
sense of our definition which requires absolute convergence of the improper
Riemann integral involved.

Exercises
5.2.1 If X is an r.v. distributed as B(n, p), calculate the kth factorial moment
E[X(X 1) (X k + 1)].
5.2.2 An honest coin is tossed independently n times and let X be the r.v.
denoting the number of Hs that occur.
iii) Calculate E(X/n), 2(X/n);
iii) If n = 100, find a lower bound for the probability that the observed
frequency X/n does not differ from 0.5 by more than 0.1;
iii) Determine the smallest value of n for which the probability that X/n does
not differ from 0.5 by more 0.1 is at least 0.95;
iv) If n = 50 and P(|(X/n) 0.5| < c) 0.9, determine the constant c. (Hint: In
(ii)(iv), utilize Tchebichevs inequality.)
5.2.3 Refer to Exercise 3.2.16 in Chapter 3 and suppose that 100 people are
chosen at random. Find the expected number of people with blood of each one
of the four types and the variance about these numbers.
5.2.4 If X is an r.v. distributed as P(), calculate the kth factorial moment
E[X(X 1) (X k + 1)].
5.2.5 Refer to Exercise 3.2.7 in Chapter 3 and find the expected number of
particles to reach the portion of space under consideration there during time
t and the variance about this number.
5.2.6 If X is an r.v. with a Hypergeometric distribution, use an approach
similar to the one used in the Binomial example in order to show that

(
mnr m + n r ) .
EX =
mr
, ( )
2 X =
m+n
(m + n) (m + n 1)
2

5.2.7 Let X be an r.v. distributed as Negative Binomial with parameters r


and p.
ii) By working as in the Binomial example, show that EX = rq/p, 2(X) =
rq/p2;
ii) Use (i) in order to show that EX = q/p and 2(X) = q/p2, if X has the
Geometric distribution.
5.2.8 Let f be the Gamma density with parameters = n, = 1. Then show
that
120 5 Moments of Random VariablesSome Moment and Probability Inequalities

n 1
x
f ( x)dx =

e .
x= 0 x!
Conclude that in this case, one may utilize the Incomplete Gamma tables (see,
for example, Tables of the Incomplete -Function, Cambridge University
Press, 1957, Karl Paerson, editor) in order to evaluate the d.f. of a Poisson
distribution at the points j = 1, 2, . . . .
5.2.9 Refer to Exercise 3.3.7 in Chapter 3 and suppose that each TV tube
costs $7 and that it sells for $11. Suppose further that the manufacturer sells an
item on money-back guarantee terms if the lifetime of the tube is less than c.
ii) Express his expected gain (or loss) in terms of c and ;
ii) For what value of c will he break even?
5.2.10 Refer to Exercise 4.1.12 in Chapter 4 and suppose that each bulb costs
30 cents and sells for 50 cents. Furthermore, suppose that a bulb is sold under
the following terms: The entire amount is refunded if its lifetime is <1,000 and
50% of the amount is refunded if its lifetime is <2,000. Compute the expected
gain (or loss) of the dealer.
5.2.11 If X is an r.v. having the Beta distribution with parameters and ,
then
ii) Show that

EX n =
( ) ( ),
+ +n
n = 1, 2, ;
( ) ( + + n)
ii) Use (i) in order to find EX and 2(X).
5.2.12 Let X be an r.v. distributed as Cauchy with parameters and 2. Then
show that E|X| = .
5.2.13 If the r.v. X is distributed as Lognormal with parameters and ,
compute EX, 2(X).
5.2.14 Suppose that the average monthly water consumption by the resi-
dents of a certain community follows the Lognormal distribution with = 104
cubic feet and = 103 cubic feet monthly. Compute the proportion of the
residents who consume more than 15 103 cubic feet monthly.
5.2.15 Let X be an r.v. with finite third moment and set = EX, 2 = 2(X).
Define the (dimensionless quantity, pure number) 1 by
3
X
1 = E .

1 is called the skewness of the distribution of the r.v. X and is a measure of
asymmetry of the distribution. If 1 > 0, the distribution is said to be skewed to
5.1 Moments of Random Exercises
Variables 121

the right and if 1 < 0, the distribution is said to be skewed to the left. Then show
that:
iii) If the p.d.f. of X is symmetric about , then 1 = 0;
iii) The Binomial distribution B(n, p) is skewed to the right for p < 1
2
and is
skewed to the left for p > 12 ;
iii) The Poisson distribution P() and the Negative Exponential distribution
are always skewed to the right.
5.2.16 Let X be an r.v. with EX4 < and define the (pure number) 2 by
4
X
2 = E

3, where = EX , 2 = 2 X . ( )
2 is called the kurtosis of the distribution of the r.v. X and is a measure of
peakedness of this distribution, where the N(0, 1) p.d.f. is a measure of
reference. If 2 > 0, the distribution is called leptokurtic and if 2 < 0, the
distribution is called platykurtic. Then show that:
ii) 2 < 0 if X is distributed as U(, );
ii) 2 > 0 if X has the Double Exponential distribution (see Exercise 3.3.13(iii)
in Chapter 3).
5.2.17 Let X be an r.v. taking on the values j with probability pj = P(X = j),
j = 0, 1, . . . . Set

()
G t = pj t j ,
j =0
1 t 1.

The function G is called the generating function of the sequence {pj}, j 0.


iii) Show that if |EX| < , then EX = d/dt G(t)|t = 1;
iii) Also show that if |E[X(X 1) (X k + 1)]| < , then

[ ( ) (
E X X 1 X k +1 = )] dk
dt k
Gt () t =1 ;

iii) Find the generating function of the sequences

n j n j
p q , j 0, 0 < p < 1, q = 1 p
j

and

j
e , j 0, > 0;
j!
122 5 Moments of Random VariablesSome Moment and Probability Inequalities

iv) Utilize (ii) and (iii) in order to calculate the kth factorial moments of X
being B(n, p) and X being P(). Compare the results with those found in
Exercises 5.2.1 and 5.2.4, respectively.

5.3 Conditional Moments of Random Variables


If, in the preceding definitions, the p.d.f. f of the r. vector X is replaced by a
conditional p.d.f. f(xj , . . . , xj |xi , . . . , xi ), the resulting moments are called
1 n 1 m

conditional moments, and they are functions of xi , . . . , xi . 1 m

Thus

x f x x ( )
(
)
2 2 1
E X 2 X1 = x1 = x 2

x f x x dx ,
2 2 1 2 ( )

[ ( )] f (x x )
2

x 2 E X 2 X1 = x1
( )
2 1
2 X 2 X1 = x1 = x 2


[ ( )] f (x x )dx .
2

x 2 E X 2 X1 = x1 2 1 2

For example, if (X1, X2) has the Bivariate Normal distribution, then f(x2 |x1)
is the p.d.f. of an N(b, 22 (1 2)) r.v., where

2
b = 2 +
1
(
x1 1 . )
Hence

(
E X 2 X1 = x1 = 2 + ) 2
1
(x1 1 . )
Similarly,

(
E X1 X 2 = x 2 = 1 + ) 1
2
(
x2 2 . )
Let X1, X2 be two r.v.s with joint p.d.f f(x1, x2). We just gave the definition
of E(X2|X1 = x1) for all x1 for which f(x2|x1) is defined; that is, for all x1 for which
fX (x1) > 0. Then E(X2|X1 = x1) is a function of x1. Replacing x1 by X1 and writing
1

E(X2|X1) instead of E(X2|X1 = x1), we then have that E(X2|X1 ) is itself an r.v.,
and a function of X1. Then we may talk about the E[E(X2|X1)]. In connection
with this, we have the following properties:
5.3 Conditional
5.1 Moments of Random Variables 123

5.3.1 Some Basic Properties of the Conditional Expectation


(CE1) If E(X2) and E(X2|X1) exist, then E[E(X2|X1)] = E(X2) (that is,
the expectation of the conditional expectation of an r.v. is the
same as the (unconditional) expectation of the r.v. in question).
It suffices to establish the property for the continuous case only, for the
proof for the discrete case is quite analogous. We have

[( )] (
E E X 2 X1 = x 2 f x 2 x1 dx 2 fX x1 dx1


) 1
( )
( ) ( )

= x 2 f x 2 x1 fX x1 dx 2 dx1
1

( ) ( )

= x 2 f x1 , x 2 dx 2 dx1 = x 2 f x1 , x 2 dx1 dx 2

( )
= x 2 f x1 , x 2 dx1 dx 2 = x 2 fX x 2 dx 2 = E X 2 . ( ) ( )

REMARK 7 Note that here all interchanges of order of integration are legiti-
mate because of the absolute convergence of the integrals involved.
(CE2) Let X1, X2 be two r.v.s, g(X1) be a (measurable) function of X1
and let that E(X2) exists. Then for all x1 for which the
conditional expectations below exist, we have

[ ( ) ] ( )(
E X 2 g X1 X1 = x1 = g x1 E X 2 X1 = x1 )
or

[ ( ) ] ( )(
E X 2 g X1 X1 = g X1 E X 2 X1 . )
Again, restricting ourselves to the continuous case, we have

[ ( ) ] ( )( ) ( ) ( )

E X 2 g X1 X1 = x1 = x 2 g x1 f x 2 x1 dx 2 = g x1 x 2 f x 2 x1 dx 2

( ) (
= g x1 E X 2 X1 = x1 . )
In particular, by taking X2 = 1, we get
(CE2) For all x1 for which the conditional expectations below exist, we
have E[g(X1)|X1 = x1] = g(x1) (or E[g(X1)|X1] = g(X1)).
(CV) Provided the quantities which appear below exist, we have

[(
2 E X 2 X1 2 X 2)] ( )
and the inequality is strict, unless X2 is a function of X1 (on a set of
probability one).

Set ( ) ( )
= E X 2 , X1 = E X 2 X1 . ( )
124 5 Moments of Random VariablesSome Moment and Probability Inequalities

Then

( ) ( ) = E{[X (X )] + [ (X ) ]}
2 2
2 X2 = E X2 2 1 1

= E[ X ( X )] + E[ ( X ) ] + 2 E{[ X ( X )][ ( X ) ]}.


2 2
2 1 1 2 1 1

Next,

{[ ( )][ ( ) ]}
E X 2 X1 X1

= E[ X ( X )] E[ ( X )] E( X ) + E[ ( X )]
2 1
2
1 2 1

= E{E[ X ( X ) X ]} E[ ( X )] E[ E( X X )]
2 1 1
2
1 2 1

+ E[ ( X )] (by (CE1)),
1

and this is equal to

[ ( )] [ ( )]
E 2 X1 E 2 X1 E X1 + E X1 [ ( )] [ ( )] (by (CE2)),
which is 0. Therefore

( ) [ ( )] [( ) ]
2 2
2 X 2 = E X 2 X1 + E X1 ,

and since

[ ( )]
2
E X 2 X1 0,

we have

( ) [( )
2 X 2 E X1 2 = 2 E X 2 X1 . ] [( )]
The inequality is strict unless

[ ( )]
2
E X 2 X1 = 0.

But

[ ( )] [ ( )] [ ( )]
2
E X 2 X1 = 2 X 2 X1 , since E X 2 X1 = = 0.

Thus 2[X2 (X1)] = 0 and therefore X2 = (X1) with probability one, by


Remark 8, which follows.

Exercises
5.3.1 Establish properties (CEI) and (CE2) for the discrete case.
5.3.2 Let the r.v.s X, Y be jointly distributed with p.d.f. given by
5.4 Some Important Applications: Probability
5.1 Moments andof
Moment
Random
Inequalities
Variables 125

(
f x, y = ) 2
(
n n+1 )
if y = 1, . . . , x; x = 1, . . . , n, and 0 otherwise. Compute the following
(
quantities: E(X|Y = y), E(Y|X = x). Hint: Recall that nx =1 x = ( 2 ) , and
n n+1

x =1 x =
n 2 n ( )(
n + 1
6 )
2 n + 1)
.
5.3.3 Let X, Y be r.v.s with p.d.f. f given by f(x, y) = (x + y)I(0,1)(0,1)(x, y).
Calculate the following quantities: EX, 2(X), EY, 2(Y), E(X|Y = y),
2(X|Y = y).
5.3.4 Let X, Y be r.v.s with p.d.f. f given by f(x, y) = 2e(x+y)I(0,)(0,)(x, y).
Calculate the following quantities: EX, 2(X), EY, 2(Y), E(X|Y = y),
2(X|Y = y).
5.3.5 Let X be an r.v. with finite EX. Then for any r.v. Y, show that
E[E(X|Y)] = EX. (Assume the existence of all p.d.f.s needed.)
5.3.6 Consider the r.v.s X, Y and let h, g be (measurable) functions on ! into
itself such that E[h(X)g(Y)] and Eg(X) exist. Then show that

[( )( )
E h X gY X=x =h x E gY X=x. ] ()[( ) ]

5.4 Some Important Applications: Probability and Moment Inequalities


THEOREM 1 Let X be a k-dimensional r. vector and g 0 be a real-valued (measurable)
function defined on !k, so that g(X) is an r.v., and let c > 0. Then

[ ( ) ] [ c ].
E g ( X)
Pg X c

PROOF Assume X is continuous with p.d.f. f. Then

[ ( )] ( )( )

E g X = g x1 , , xk f x1 , , xk dx1 dxk

( )( ) (
= g x1 , , xk f x1 , , xk dx1 dxk + c g x1 , , xk
A A
)
f ( x , , x )dx
1 k 1 dxk ,

where A = {(x1, . . . , xk) !k; g(x1, . . . , xk) c}. Then


126 5 Moments of Random VariablesSome Moment and Probability Inequalities

[ ( )] ( )( )
E g X g x1 , , xk f x1 , , xk dx1 dxk
A

c f ( x , , x )dx dx
1 k 1 k
A

= cP[ g(X ) A] = cP[ g(X ) c].


Hence P[g(X) c] E[g(X)]/c. The proof is completely analogous if X is of the
discrete type; all one has to do is to replace integrals by summation signs.

5.4.1 Special Cases of Theorem 1


1. Let X be an r.v. and take g(X) = |X |r, = E(X), r > 0. Then
r
EX
[ ]
P X c = P X c r
r
.
cr

This is known as Markovs inequality.


2. In Markovs inequality replace r by 2 to obtain

( )=
2
EX 2 X
[ ]
2

P X c = P X c 2
2
= .
c2 c 2
c 2

This is known as Tchebichevs inequality. In particular, if c = k, then

[
P X k ] k
1
2
k
1
[
; equivalently, P X < k 1 2 . ]
REMARK 8 Let X be an r.v. with mean and variance 2 = 0. Then
Tchebichevs inequality gives: P[|X | c] = 0 for every c > 0. This result and
Theorem 2, Chapter 2, imply then that P(X = ) = 1 (see also Exercise 5.4.6).
LEMMA 1 Let X and Y be r.v.s such that

( )
E X = E Y = 0, ( ) ( )
2 X = 2 Y = 1.( )
Then

( )
E 2 XY 1 or, equivalently, 1 E XY 1, ( )
and

( )
E XY = 1 if any only if ( )
P Y = X = 1,
E( XY ) = 1 if any only if P(Y = X ) = 1.

PROOF We have

(
0 E X Y )
2
(
= E X 2 2 XY + Y 2 )
( )
= EX 2 2 E XY + EY 2 = 2 2 E XY ( )
5.4 Some Important Applications: Probability
5.1 Moments andof
Moment
Random
Inequalities
Variables 127

and

0 E X +Y( )
2
(
= E X 2 + 2 XY + Y 2 )
( )
= EX + 2 E XY + EY 2 = 2 + 2 E XY .
2
( )
Hence E(XY) 1 and 1 E(XY), so that 1 E(XY) 1. Now let P(Y = X)
= 1. Then E(XY) = EY2 = 1, and if P(Y = X) = 1, then E(XY) = EY2 = 1.
Conversely, let E(XY) = 1. Then

( ) ( ) [E(X Y )] = E(X Y )
2 2 2
2 X Y = E X Y
= EX 2 2 E( XY ) + EY = 1 2 + 1 = 0,
2

so that P(X = Y) = 1 by Remark 8; that is, P(X = Y) = 1. Finally, let


E(XY) = 1. Then 2(X + Y) = 2 + 2E(XY) = 2 2 = 0, so that

(
P X = Y = 1. )
THEOREM 2 (CauchySchwarz inequality) Let X and Y be two random variables with
means 1, 2 and (positive) variances 12 , 22 , respectively. Then

[( )(
E 2 X 1 Y 2 12 22 , )]
or, equivalently,

[(
1 2 E X 1 Y 2 1 2 , )( )]
and

[(
E X 1 Y 2 )( )] = 1 2

if and only if

P Y = 2 + 2 X 1 = 1

( )
1
and

[(
E X 1 Y 2 )( )] = 1 2

if and only if

P Y = 2 2 X 1 = 1.
1
( )

PROOF Set

X 1 Y 2
X1 = , Y1 = .
1 2
128 5 Moments of Random VariablesSome Moment and Probability Inequalities

Then X1, Y1 are as in the previous lemma, and hence

(
E 2 X1Y1 1 )
if and only if
1 E X1Y1 1( )
becomes

[(
E 2 X 1 Y 2 )( )] 1

1
2 2
2

if and only if

[( )(
1 2 E X 1 Y 2 1 2 . )]
The second half of the conclusion follows similarly and will be left as an
exercise (see Exercise 5.4.6).
REMARK 9 A more familiar form of the CauchySchwarz inequality is
E2(XY) (EX2)(EY2). This is established as follows: Since the inequality is
trivially true if either one of EX2, EY2 is , suppose that they are both finite
and set Z = X Y, where is a real number. Then 0 EZ2 = (EX2)2
2[E(XY)] + EY2 for all , which happens if and only if E2(XY) (EX2)(EY2)
0 (by the discriminant test for quadratic equations), or E2(XY) (EX2)(EY2).

Exercises
5.4.1 Establish Theorem 1 for the discrete case.
5.4.2 Let g be a (measurable) function defined on ! into (0, ). Then, for any
r.v. X and any > 0,

Eg X( ).
[( ) ]
P g X

If furthermore g is even (that is, g(x) = g(x)) and nondecreasing for x 0, then

Eg X ( ).
(
P X ) g ( )

5.4.3 For an r.v. X with EX = and 2(X) = 2, both finite, use Tchebichevs
inequality in order to find a lower bound for the probability P(|X | < k).
Compare the lower bounds for k = 1, 2, 3 with the respective probabilities
when X N(, 2).
5.5 Covariance, Correlation
5.1 Coefficient
Moments of and
Random
Its Interpretation
Variables 129

5.4.4 Let X be an r.v. distributed as 40 2


. Use Tchebichevs inequality
in order to find a lower bound for the probability P(|(X/40) 1| 0.5),
and compare this bound with the exact value found from Table 3 in Appendix
III.
5.4.5 Refer to Remark 8 and show that if X is an r.v. with EX = (finite) such
that P(|X | c) = 0 for every c > 0, then P(X = ) = 1.
5.4.6 Prove the second conclusion of Theorem 2.
5.4.7 For any r.v. X, use the CauchySchwarz inequality in order to show
that E|X| E1/2X2.

5.5 Covariance, Correlation Coefficient and Its Interpretation


In this section, we introduce the concepts of covariance and correlation coef-
ficient of two r.v.s and provide an interpretation for the latter. To this end,
for two r.v.s X and Y with means 1, 2, the (1, 1)-joint central mean, that is,
E[(X 1)(Y 2)], is called the covariance of X, Y and is denoted by
Cov(X,Y). If 1, 2 are the standard deviations of X and Y, which are assumed
to be positive, then the covariance of (X 1)/1, (Y 2)/2 is called the
correlation coefficient of X, Y and is denoted by (X, Y) or X,Y or 12 or just
if no confusion is possible; that is,

[( )(
X 1 Y 2 E X 1 Y 2
= E =
)] = Cov( X , Y )
1 2 1 2 1 2

=
( )
E XY 1 2
.
1 2

From the CauchySchwarz inequality, we have that 2 1; that is 1 1,


and = 1 if and only if
2
Y = 2 +
1
(
X 1 )
with probability 1, and = 1 if and only if

2
Y = 2
1
(
X 1 )
with probability 1. So = 1 means X and Y are linearly related. From this
stems the significance of as a measure of linear dependence between X and
Y. (See Fig. 5.1.) If = 0, we say that X and Y are uncorrelated, while if = 1,
we say that X and Y are completely correlated (positively if = 1, negatively if
= 1).
130 5 Moments of Random VariablesSome Moment and Probability Inequalities

Y
$2
#2 , $1 #1 $2
y ! #2 , $1 (x " # 1 )
#2
$2
y ! #2 " $1 (x " # 1 )
$2
#2 " $1 #1

X
0 #1
Figure 5.1

For 1 < < 1, 0, we say that X and Y are correlated (positively if > 0,
negatively if < 0). Positive values of may indicate that there is a tendency of
large values of Y to correspond to large values of X and small values of Y to
correspond to small values of X. Negative values of may indicate that small
values of Y correspond to large values of X and large values of Y to small
values of X. Values of close to zero may also indicate that these tendencies
are weak, while values of close to 1 may indicate that the tendencies are
strong.
The following elaboration sheds more light on the intuitive interpreta-
tion of the correlation coefficient (= (X, Y)) as a measure of co-linearity
of the r.v.s X and Y. To this end, for > 0, consider the line y = 2 + ( x 1 ) 2
1

in the xy-plane and let D be the distance of the (random) point (X, Y)
from the above line. Recalling that the distance of the point (x0, y0) from the
line ax + by + c = 0 is given by ax0 + by0 + c a 2 + b 2 . we have in the present
case:
1 12
D= X Y + 1 2 1 1+ ,
2 2 22

1
since here a = 1, b = and c = 1 2 1 . Thus,
2 2
2
12
D = X 1 Y + 1 2 1
2
1 + ,
2 2 22

and we wish to evaluate the expected squared distance of (X, Y) from the line
y = 2 + ( x 1 ) ; that is, ED2. Carrying out the calculations, we find
2
1

( 2
1 ) (
+ 22 D 2 = 22 X 2 + 12Y 2 2 1 2 XY + 2 2 1 2 2 1 X )
( ) ( )
2
2 1 1 2 2 1 Y + 1 2 2 1 . (1)

Taking the expectations of both sides in (1) and recalling that


5.5 Covariance, Correlation
5.1 Coefficient
Moments of and
Random
Its Interpretation
Variables 131

EX 2 = 12 + 12 , EY 2 = 22 + 22 ( )
and E XY = 1 2 + 1 2 ,

we obtain

2 12 22
ED2 =
12 + 22
(1 ) ( > 0). (2)

Working likewise for the case that < 0, we get

2 12 22
ED2 =
12 + 22
(1+ ) ( < 0). (3)

For = 1 or = 1, we already know that (X, Y) lies on the line y =


2 + ( x 1 ) or y = 2 ( x 1 ), respectively (with probability 1). There-
2
1
2
1

fore, regardless of the value of , by the observation just made, relations (2)
and (3) are summarized by the expression

ED2 =
2 12 22
(
12 + 22
)
1 . (4)

At this point, exploiting the interpretation of an expectation as an average,


relation (4) indicates the following: For > 0, the pairs (X, Y) tend to be
arranged along the line y = 2 + ( x 1 ) . These points get closer and closer
2
1

to this line as gets closer to 1, and lie on the line for = 1. For < 0, the
pairs (X, Y) tend to be arranged along the line y = 2 ( x 1 ) . These
2
1

points get closer and closer to this line as gets closer to 1, and lie on
this line for = 1. For = 0, the expected distance is constantly equal to
2 12 22 /( 12 + 22 ) from either one of the lines y = 2 + ( x 1 ) and
2

y = 2 ( x 1 ) , which is equivalent to saying that the pairs (X, Y) may lie


1
2
1

anywhere in the xy-plane. It is in this sense that is a measure of co-linearity


of the r.v.s X and Y.
The preceding discussion is based on the paper A Direct Development
of the Correlation Coefficient by Leo Katz, published in the American
Statistician, Vol. 29 (1975), page 170. His approach is somewhat different
and is outlined below. First, consider the r.v.s X1 and Y1 as defined in
the proof of Theorem 2; unlike the original r.v.s X and Y, the normalized
r.v.s X1 and Y1 are dimensionless. Through the transformations x1 = x 1
1

and y1 = y , we move from the xy-plane to the x1y1-plane. In this latter plane,
2
2

look at the point (X1, Y1) and seek the line Ax1 + By1 + C = 0
from which the expected squared distance of (X1, Y1) is minimum. That
is, determine the coefficients A, B and C, so that ED12 is minimum, where
132 5 Moments of Random VariablesSome Moment and Probability Inequalities

D1 = AX 1 + BY1 + C1 A2 + B2 . Expanding D12, taking expectations, and


noticing that

EX1 = EY1 = 0, EX12 = EY12 = 1, and E X1Y1 = , ( )


we obtain

2 AB C2
ED12 = 1 + + . (5)
A2 + B2 A2 + B2

Clearly, for ED12 to be minimized it is necessary that C = 0. Then, by (5), the


expression to be minimized is

2 AB
ED12 = 1 + . (6)
A2 + B2

At this point, observe that

(
A+ B )
2
( ) ( )
= A2 + B2 2 AB 0 A2 + B2 2 AB = A B , ( )
2

or equivalently,

2 AB
1 1. (7)
A2 + B2

From (6) and (7), we conclude that:


If > 0, ED12 is minimized for A22AB+ B2
= 1 and the minimum is 1 .
If < 0, ED12 is minimized for A22AB+ B2
= 1 and the minimum is 1 + .
Finally, if = 0, the ED1 is constantly equal to 1; there is no minimizing
2

line (through the origin) Ax1 + By1 = 0. However, A22AB + B2


= 1 if and only if A =
B, and A22AB
+ B2
= 1 if and only if A = B. The corresponding lines are y1 = x1, the
main diagonal, and y1 = x1. Also observe that both minima of ED12 (for > 0
and < 0), and its constant value 1 (for = 0) are expressed by a single form,
namely, 1 ||.
To summarize: For > 0, the ED12 is minimized for the line y1 = x1; for
< 0, the ED12 is minimized for the line y1 = x1; for = 0, ED12 = 1, there is
no minimizing line. From this point on, the interpretation of as a measure of
co-linearity (of X1 and Y1) is argued as above, with the lines y = 2 + ( x 1 )
2

and y = 2 ( x 1 ) being replaced by the lines y1 = x1 and y1 = x1,


1
2
1

respectively.
5.1 Moments of Random Exercises
Variables 133

Exercises
5.5.1 Let X be an r.v. taking on the values 2, 1, 1, 2 each with probability
1 . Set Y = X2 and compute the following quantities: EX, 2(X), EY, 2(Y),
4
(X, Y).
5.5.2 Go through the details required in establishing relations (2), (3) and
(4).
5.5.3 Do likewise in establishing relation (5).
5.5.4 Refer to Exercise 5.3.2 (including the hint given there) and
iii) Calculate the covariance and the correlation coefficient of the r.v.s X and
Y;
iii) Referring to relation (4), calculate the expected squared distance of (X,
Y) from the appropriate line y = 2 + ( x 1 ) or y = 2 ( x 1 )
2
1
2
1

(which one?);
iii) What is the minimum expected squared distance of (X1, Y1) from
the appropriate line y = x or y = x (which one?) where X 1 = X and 1
1

Y1 = Y .
2
2


( ) .
2
n n n+1
Hint: Recall that
x = 2
3



x =1

5.5.5 Refer to Exercise 5.3.2 and calculate the covariance and the correla-
tion coefficient of the r.v.s X and Y.
5.5.6 Do the same in reference to Exercise 5.3.3.
5.5.7 Repeat the same in reference to Exercise 5.3.4.

5.5.8 Show that (aX + b, cY + d) = sgn(ac)(X, Y), where a, b, c, d are


constants and sgn x is 1 if x > 0 and is 1 if x < 0.
5.5.9 Let X and Y be r.v.s representing temperatures in two localities, A and
B, say, given in the Celsius scale, and let U and V be the respective tempera-
tures in the Fahrenheit scale. Then it is known that U and X are related as
follows: U = 59 X + 32, and likewise for V and Y. Fit this example in the model
of Exercise 5.5.8, and conclude that the correlation coefficients of X, Y and U,
V are identical, as one would expect.
5.5.10 Consider the jointly distributed r.v.s X, Y with finite second moments
and 2(X) > 0. Then show that the values and for which E[Y (X + )]2
is minimized are given by
134 5 Moments of Random VariablesSome Moment and Probability Inequalities

Y( ) X, Y .
= EY EX , =
(X )
( )
(The r.v. Y = X + is called the best linear predictor or Y, given X.)

5.5.11 If the r.v.s X1 and X2 have the Bivariate Normal distribution, show
that the parameter is, actually, the correlation coefficient of X1 and X2. (Hint:
Observe that the exponent in the joint p.d.f. of X1 and X2 may be written as
follows:

1 x 2 x 1 1 x 2 2 x 2 2
2

1 1
2 +
(
2 1 2 ) 1


1 2 2

(x ) (x )
2 2
1 b 2
=
1

2 2
+
2
2
, where b = 2 +
1
( )
x 1 1 .
1
2 2 1 2

This facilitates greatly the integration in calculating E(X1X2).

5.5.12 If the r.v.s X1 and X2 have jointly the Bivariate Normal distribution
with parameters 1, 2, 12 , 22 and , calculate E(c1X1 + c2X2) and 2(c1X1 +
c2X2) in terms of the parameters involved, where c1 and c2 are real constants.

5.5.13 For any two r.v.s X and Y, set U = X + Y and V = X Y. Then

iii) Show that P(UV < 0) = P(|X| < |Y|);


iii) If EX2 = EY2 < , then show that E(UV) = 0;
iii) If EX2, EY2 < and 2(X) = 2(Y), then U and V are uncorrelated.

5.5.14 If the r.v.s Xi, i = 1, . . . , m and Yj, j = 1, . . . , n have finite second


moments, show that

m n m n
Cov X i , Yj = Cov X i , Yj .
i =1 j =1 i =1 j =1
( )

5.6* Justification of Relation (2) in Chapter 2


As a final application of the results of this chapter, we give a general proof of
Theorem 9, Chapter 2. To do this we remind the reader of the definition of the
concept of the indicator function of a set A.
5.6* Justification
5.1 Moments
of Relation
of Random
(2) in Chapter
Variables
2 135

Let A be an event in the sample space S. Then the indicator function of A,


denoted by IA, is a function on S defined as follows:

1 if s A
()
IA s =
0 if s A .
c

The following are simple consequences of the indicator function:

n
I I Aj = I A
n
j=1 j
(8)
j =1

n
I Aj = I A , (9)
n
j=1
j =1
j

and, in particular,

IA = 1 IA.
c (10)

Clearly,

( )
E IA = P A ( ) (11)

and for any X1, . . . , Xr, we have

(1 X )(1 X ) (1 X ) = 1 H ( )
r
1 2 r 1 + H 2 + 1 H r , (12)

where Hj stands for the sum of the products Xi Xi , where the summation 1 j

extends over all subsets {i1, i2, . . . , ij} of the set {1, 2, . . . , r}, j = 1, . . . , r. Let
, be such that: 0 < , and + r. Then the following is true:

+
Xi
j
1
X i H J = ( )

H + , (13)

where J = {i1, . . . , i} is the typical member of all subsets of size of the set
{1, 2, . . . , r}, H(J) is the sum of the products Xj Xj , where the summation
1

extends over all subsets of size of the set {1, . . . , r} J, and J is meant to

extend over all subsets J of size of the set {1, 2, . . . , r}.


The justification of (13) is as follows: In forming H+, we select ( + ) Xs
from the available r Xs in all possible ways, which is ( +r ). On the other hand,
for each choice of J, there are ( r ) ways of choosing Xs from the remaining
(r ) Xs. Since there are (r ) choices of J, we get (r )( r ) groups (products)
of ( + ) Xs out of r Xs. The number of different groups of ( + ) Xs out
136 5 Moments of Random VariablesSome Moment and Probability Inequalities

of r Xs is ( +r ). Thus among the (r )( r ) groups of ( + ) Xs out of r Xs, the


number of distinct ones is given by

r r r! r ! ( )

(
! r ! ! r !
=
) ( )
r r!

+
+ ! (
r ! )( )
=
( + )! = + .
! !

This justifies (13).


Now clearly,
Bm = Ai Ai Aic Aic ,
1 m m +1 M
Jm

where the summation extends over all choices of subsets Jm = {i1, . . . , im} of the
set {1, 2, . . . , M} and Bm is the one used in Theorem 9, Chapter 2. Hence

IB = IA
m i1 Aim Aicm + 1 AicM
Jm

= IA IA 1 IA
Jm
i1 im ( im + 1 ) (1 I ) (by (8), (9), (10))
AiM

( )
= I A I A 1 H1 J m + H 2 J m + 1 ( ) ( ) ( )
H M m J m
M m

J m
i1
im

(by (12)).
Since

IA
Jm
i1 im
m + k
I A H k Jm = ( )
m
H m+ k (by (13)),
we have
m + 1 m + 2 M
( )
M m
I B = Hm H m+1 + H m+ 2 + 1 HM .
m
m m m
Taking expectations of both sides, we get (from (11) and the definition of Sr in
Theorem 9, Chapter 2)

m + 1 m + 2 M
( ) ( )
M m
P Bm = S m S m+1 + S m+ 2 + 1 SM ,
m m m
as was to be proved.
5.6* Justification
5.1 Moments
of Relation
of Random
(2) in Chapter
Variables
2 137

(For the proof just completed, also see pp. 8085 in E. Parzens book
Modern Probability Theory and Its Applications published by Wiley, 1960.)
REMARK 10 In measure theory the quantity IA is sometimes called the char-
acteristic function of the set A and is usually denoted by A. In probability
theory the term characteristic function is reserved for a different concept and
will be a major topic of the next chapter.
138 6 Characteristic Functions, Moment Generating Functions and Related Theorems

Chapter 6

Characteristic Functions,
Moment Generating Functions
and Related Theorems

6.1 Preliminaries
The main subject matter of this chapter is the introduction of the concept of
the characteristic function of an r.v. and the discussion of its main properties.
The characteristic function is a powerful mathematical tool, which is used
profitably for probabilistic purposes, such as producing the moments of an r.v.,
recovering its distribution, establishing limit theorems, etc. To this end, recall
that for z !, eiz = cos z + i sin z, i = 1 , and in what follows, i may
be treated formally as a real number, subject to its usual properties: i2 = 1,
i3 = i, i 4 = 1, i 5 = i, etc.
The sequence of lemmas below will be used to justify the theorems which
follow, as well as in other cases in subsequent chapters. A brief justification for
some of them is also presented, and relevant references are given at the end of
this section.

LEMMA A Let g1, g2 : {x1, x2, . . .} [0, ) be such that

( ) ( )
g1 x j g 2 x j , j = 1, 2, ,

and that xj g2(xj) < . Then xj g1(xj) < .

PROOF If the summations are finite, the result is immediate; if not, it follows
by taking the limits of partial sums, which satisfy the inequality. "
b
LEMMA A Let g1, g2 : ! [0, ) be such that g1(x) g2(x), x !, and that a g1 (x)dx

exists for every a, b, ! with a < b, and that g 2 (x)dx < . Then g1(x)dx
< .

PROOF Same as above replacing sums by integrals. "

138
6.5 The Moment Generating
6.1 Preliminaries
Function 139

LEMMA B Let g : {x1, x2, . . .} ! and xj|g(xj)| < . Then xj g(xj) also converges.
PROOF The result is immediate for finite sums, and it follows by taking the
limits of partial sums, which satisfy the inequality. "
b
LEMMA B Let g : ! ! be such that a g(x)dx exists for every a, b, ! with a < b, and that

|g( x ) |dx < . Then g(x)dx also converges.
PROOF Same as above replacing sums by integrals. "
The following lemma provides conditions under which the operations
of taking limits and expectations can be interchanged. In more advanced
probability courses this result is known as the Dominated Convergence
Theorem.
LEMMA C Let {Xn}, n = 1, 2, . . . , be a sequence of r.v.s, and let Y, X be r.v.s such
that |Xn(s)| Y(s), s S, n = 1, 2, . . . and Xn(s) X(s) (on a set of ss of
probability 1) and E(Y) < . Then E(X) exists and E(Xn) n
E(X), or
equivalently,

( ) (
lim E X n = E lim X n .
n n
)
REMARK 1 The index n can be replaced by a continuous variable.
The next lemma gives conditions under which the operations of differen-
tiation and taking expectations commute.
LEMMA D For each t T (where T is ! or an appropriate subset of it, such as the interval
[a, b]), let X(; t) be an r.v. such that (t)X(s; t) exists for each s S and
t T. Furthermore, suppose there exists an r.v. Y with E(Y) < and such
that


t
( )
X s; t Y s ,() s S , t T .

Then

d
dt [ ( )]
( )
E X ; t = E X ; t , for all t T .
t
The proofs of the above lemmas can be found in any book on real vari-
ables theory, although the last two will be stated in terms of weighting func-
tions rather than expectations; for example, see Advanced Calculus, Theorem
2, p. 285, Theorem 7, p. 292, by D. V. Widder, Prentice-Hall, 1947; Real
Analysis, Theorem 7.1, p. 146, by E. J. McShane and T. A. Botts, Van
Nostrand, 1959; The Theory of Lebesgue Measure and Integration, pp. 6667,
by S. Hartman and J. Mikusinski, Pergamon Press, 1961. Also Mathematical
Methods of Statistics, pp. 4546 and pp. 6668, by H. Cramr, Princeton
University Press, 1961.
140 6 Characteristic Functions, Moment Generating Functions and Related Theorems

6.2 Definitions and Basic TheoremsThe One-Dimensional Case


Let X be an r.v. with p.d.f. f. Then the characteristic function of X (ch.f. of X),
denoted by X (or just when no confusion is possible) is a function defined on
!, taking complex values, in general, and defined as follows:

() [ ( ) ( ) ( ) ( )]
e itx f x = cos tx f x + i sin tx f x
()
X t = E e itX [ ]
= x x

() [ ( ) ( ) ( ) ( )]

e f x dx =
cos tx f x + i sin tx f x dx
itx

[ ( ) ( )] [ ( ) ( )]
cos tx f x + i sin tx f x

= x x

( )() ( )()

cos tx f x dx + i sin tx f x dx.

By Lemmas A, A, B, B, X(t) exists for all t !. The ch.f. X is also called the
Fourier transform of f.
The following theorem summarizes the basic properties of a ch.f.
THEOREM 1 (Some properties of ch.fs)
i) X(0) = 1.
ii)|X(t)| 1.
iii)X is continuous, and, in fact, uniformly continuous.
iv) X+d(t) = eitdX(t), where d is a constant.
v) cX(t) = X(ct), where c is a constant.
vi) cX+d(t) = eitdX(ct).

dn
vii)
dt n
X t () = inE(Xn), n = 1, 2, . . . , if E|Xn| < .
t =0

PROOF
i) X(t) = EeitX. Thus X(0) = Eei0X = E(1) = 1.
ii) |X(t)| = |EeitX| E|eitX| = E(1) = 1, because |eitX| = 1. (For the proof of the
inequality, see Exercise 6.2.1.)

( )
iii) X t + h X t = E e
() ( )
i t+ h X
e itX

[ ( )] ( )
= E e itX e ihX 1 E e itX e ihX 1

= E e ihX 1.

Then
6.2 Definitions and Basic TheoremsThe
6.5 The Moment
One-Dimensional
Generating Function
Case 141

h0
( ) ()
lim X t + h X t lim E e ihX 1 = E lim e ihX 1 = 0,
h0 [ h0 ]
provided we can interchange the order of lim and E, which here can be
done by Lemma C. We observe that uniformity holds since the last ex-
pression on the right is independent of t.
iv) X+d(t) = Eeit(X+d) = E(eitXeitd) = eitd EeitX = eitd X(t).
v) cX(t) = Eeit(cX) = Eei(ct)X = X(ct).
vi) Follows trivially from (iv) and (v).

n
vii)
dn
dt n X
dn
()
t = n Ee itX = E n e itX = E i n X n e itX ,
dt t
( )
provided we can interchange the order of differentiation and E. This can
be done here, by Lemma D (applied successively n times to X and its
n 1 first derivatives), since E|X n| < implies E|X k| < , k = 1, . . . , n
(see Exercise 6.2.2). Thus

dn
dt n
X t () ( )
= inE X n . "
t =0

REMARK
n
2 From part (vii) of the theorem we have that E(Xn) =
(i) n d
dt n
X(t)|t = 0, so that the ch.f. produces the nth moment of the r.v.
REMARK 3 If X is an r.v. whose values are of the form x = a + kh, where a,
h are constants, h > 0, and k runs through the integral values 0, 1, . . . , n or 0,
1, . . . , or 0, 1, . . . , n or 0, 1, . . . , then the distribution of X is called a lattice
distribution. For example, if X is distributed as B(n, p), then its values are of
the form x = a + kh with a = 0, h = 1, and k = 0, 1, . . . , n. If X is distributed as
P(), or it has the Negative Binomial distribution, then again its values are of
the same form with a = 0, h = 1, and k = 0, 1, . . . . If now is the ch.f. of X, it
can be shown that the distribution of X is a lattice distribution if and only if
|(t)| = 1 for some t 0. It can be readily seen that this is indeed the case in the
cases mentioned above (for example, (t) = 1 for t = 2). It can also be shown
that the distribution of X is a lattice distribution, if and only if the ch.f. is
periodic with period 2 (that is, (t + 2) = (t), t !).
In the following result, the ch.f. serves the purpose of recovering the
distribution of an r.v. by way of its ch.f.
THEOREM 2 (Inversion formula) Let X be an r.v. with p.d.f. f and ch.f. . Then if X is of the
discrete type, taking on the (distinct) values xj, j 1, one has

( ) 1
()
T itx j
i) f x j = lim
T 2T
T
e t dt , j 1.

If X is of the continuous type, then


142 6 Characteristic Functions, Moment Generating Functions and Related Theorems

1 e ith itx
() 1
()
T
ii) f x = lim lim
h0 T 2 T ith
e t dt


and, in particular, if | (t )|dt < , then (f is bounded and continuous and)

1 itx
ii) f(x) = e (t)dt.
2
PROOF (outline) i) The ch.f. is continuous, by Theorem 1(iii), and since
T itx
so is eitxj, it follows that the integral T e (t)dt exists for every T(> 0). We j

have then

itx
1
() 1
( )
T T
e itx f x k dt
itx j
e t dt =
e
j k

2T T 2T T
k
(
) f x dt
1
( )
T
e
it x x

2T T k
= k
k j


1 T it( x x )
= f xk
k
( )
2T T
e dt k j

(the interchange of the integral and summations is valid here). That is,

( ) dt .
1
() ( ) 21T
T T
t dt = f x k
itx j it x k x j

2T T e k
T
e (1)

But

( ) dt =
[cos t(x ) ( )]
T it x k x j T
T
e T
k x j + i sin t x k x j dt

( ) ( )
T T
= cos t x k x j dt + i sin t x k x j dt
T T

= cos t ( x x )dt , since sin z is an odd function. That is,


T
k j
T
( ) dt = cos t x
( x )dt .
T it x k x j T
T
e T
k j (2)

If xk = xj, the above integral is equal to 2T, whereas, for xk xj,

( ) 1
( )
T T
T
cos t x k x j dt =
xk x j T
d sin t x k x j

=
( )
sin T x k x j sin T x k x j [ ( )]
xk x j

=
(
2 sin T x k x j ).
xk x j
6.2 Definitions and Basic TheoremsThe
6.5 The Moment
One-Dimensional
Generating Function
Case 143

Therefore,

1, if xk = x j
( ) dt = sin T x x
1
( k j)
T it x k x j

2T
T
e
, if xk x j.
(3)

(
T x k x j )
sinT ( x x )
But k
x x
j
x 1 x , a constant independent of T, and therefore, for
k j
sin T ( x x )
k j

xk xj, lim T ( x x ) = 0 , so that


k j

T k j

1 T (
it x k x j ) dt = 1, if xk = x j
lim
T 2T e
xk x j.
(4)
0, if
T

By means of (4), relation (1) yields

1
() ( ) 21T ( )
T T
t dt = lim f xk
it xk x j
T e
itx j
lim e dt
T 2T T
k
T

( ) 1 ( )
T
= f xk lim
it xk x j

k
T 2T T e dt

(the interchange of the limit and summation is legitimate here)


( ) dt = f x ,
( ) 1
( j)
T
= f x j + lim
it x k x j

k j
T 2T T e
as was to be seen
ii) (ii) Strictly speaking, (ii) follows from (ii). We are going to omit (ii)
entirely and attempt to give a rough justification of (ii). The assumption that
itx
| (t )|dt < implies that e (t)dt exists, and is taken as follows for every
arbitrary but fixed x !:

() ()
T

e itx t dt = lim e itx t dt .
(0 < )T T
(5)

But

()
e itx t dt = e itx e ity f y dydt ()
T T
T T

() ()
it ( y x ) it ( y x )
f y dy dt = f y e dt dy,
T T
= e (6)
T T
where the interchange of the order of integration is legitimate here. Since the
integral (with respect to y) is zero over a single point, we may assume in the
sequel that y x. Then

T (
it y x ) 2 sin T y x ( ),
T e dt =
yx
(7)

as was seen in part (i). By means of (7), relation (6) yields


144 6 Characteristic Functions, Moment Generating Functions and Related Theorems

sin T y x ( ) dy.
() ()

e t dt = 2 lim f y (8)
itx
T yx
Setting T(y x) = z, expression (8) becomes

z sin z
()


e itx t dt = 2 lim f x +
T T z
dz

()
= 2 f x = 2f x , ()
by taking the

limit under the integral sign, and by using continuity of f and the
fact that sinz z dz = . Solving for f(x), we have

() 1
()

f x =
2
e itx t dt ,

as asserted. "
EXAMPLE 1 Let X be B(n, p). In the next section, it will be seen that X(t) = (peit + q)n. Let
us apply (i) to this expression. First of all, we have

1
() 1
( pe )
T T n
e itx t dt = + q e itx dt
it

2T T 2T T

n
r ( pe )q
n
1 T r
n r
= it
e itx dt
2T T
r = 0
1 n
n T ( )
r p q n r i r x t
= r
e dt
2T r= 0
T

1 n
n 1 ( )
( )
T
r p q
i r x t

n r
= r
e i r x dt
2T r= 0 i( r x ) T
r x

1 n x n x T
+ p q T dt
2T x
i ( r x )T i ( r x )T
n
n e e 1 n x n x
= p r qn r + p q 2T
r= 0 r
r x
2Ti r x 2T x ( )
n
n
= p r qn r
sin r x T n x n x
+ p q .
( )
r= 0 r
r x
rx T x ( )
Taking the limit as T , we get the desired result, namely

n
()
f x = p x q n x .
x
6.5 The Moment GeneratingExercises
Function 145

(One could also use (i) for calculating f(x), since is, clearly, periodic with
period 2.)

EXAMPLE 2 For an example of the continuous type, let X be N(0, 1). In the next section, we

will see that X(t) = et /2. Since |(t)| = et /2, we know that | (t )|dt < , so that
2 2

(ii) applies. Thus we have


1 1
( ) ()


2
f x = e itx t dt = e itx e t 2
dt
2 2

1 ( 1 2 )( t + 2 itx ) 1 ( 1 2 ) t + 2 t ( ix ) + ( ix ) ( 1 2 )( ix )
2 2 2
2

2 2
= e dt = e e dt

( 1 2 )x 2
( 1 2 )x 2

e 1 ( 1 2 )( t + ix ) e 2
1 ( 1 2 )u2

=
2

2
e dt =
2

2
e du

( )
1 2 x2
e 1 2
= 1 = e x 2
,
2 2
as was to be shown.

THEOREM 3 (Uniqueness Theorem) There is a one-to-one correspondence between the


characteristic function and the p.d.f. of a random variable.

PROOF The p.d.f. of an r.v. determines its ch.f. through the definition of the
ch.f. The converse, which is the involved part of the theorem, follows from
Theorem 2. "

Exercises
6.2.1 Show that for any r.v. X and every t !, one has |EeitX| E|eitX|(= 1).
(Hint: If z = a + ib, a, b !, recall that z = a 2 + b 2 . Also use Exercise 5.4.7
in Chapter 5 in order to conclude that (EY)2 EY2 for any r.v. Y.)

6.2.2 Write out detailed proofs for parts (iii) and (vii) of Theorem 1 and
justify the use of Lemmas C, D.

6.2.3 For any r.v. X with ch.f. X, show that X(t) = X(t), t !, where the bar
over X denotes conjugate, that is, if z = a + ib, a, b !, then z = a ib.
6.2.4 Show that the ch.f. X of an r.v. X is real if and only if the p.d.f. fX of X
is symmetric about 0 (that is, fX(x) = fX(x), x !). (Hint: If X is real, then the
conclusion is reached by means of the previous exercise and Theorem 2. If fX
is symmetric, show that fX(x) = fX(x), x !.)
146 6 Characteristic Functions, Moment Generating Functions and Related Theorems

6.2.5 Let X be an r.v. with p.d.f. f and ch.f. given by: (t) = 1 |t| if |t| 1
and (t) = 0 if |t| > 1. Use the appropriate inversion formula to find f.
6.2.6 Consider the r.v. X with ch.f. (t) = e|t|, t !, and utilize Theorem 2(ii)
in order to determine the p.d.f. of X.

6.3 The Characteristic Functions of Some Random Variables


In this section, the ch.f.s of some distributions commonly occurring will be
derived, both for illustrative purposes and for later use.

6.3.1 Discrete Case


1. Let X be B(n, p). Then X(t) = (peit + q)n. In fact,
n n
( ) ( )
n n

()
x n
X t = e itx p x q n x = pe it q n x = pe it + q ,
x= 0 x x = 0 x

Hence

d
() ( )
n 1
X t = n pe it + q ipe it = inp,
dt t=0 t=0

so that E(X ) = np. Also,

d2
() (
d it
)
n 1
X t = inp pe + q e it
dt 2 t=0
dt t= 0

( )( ) ( )
n 2 n 1
= inp n 1 pe it + q ipe it e it + pe it + q ie it
t= 0
[( ) ]
= i 2 np n 1 p + 1 = np n 1 p + 1 = E X 2 ,[( ) ] ( )
so that

( ) [( ) ]
E X 2 = np n 1 p + 1 and 2 X = E X 2 EX ( ) ( ) ( ) 2

= n 2 p 2 np 2 + np n 2 p 2 = np 1 p = npq; ( )
that is, 2(X ) = npq.
2. Let X be P(). Then X(t) = ee . In fact,
it

( )
x
e
it

x
()
X t = e itx e =e
it it

= e e e = e e
.
x= 0 x! x= 0 x!
6.3 The Characteristic 6.5
Functions
The Moment
of Some
Generating
Random Variables
Function 147

Hence

d
()
it
X t = e e
ie it = i ,
dt t=0 t= 0

so that E(X ) = . Also,

d2
dt 2
X t()
t=0
=
d
dt
(
ie e e
it
+it
) t=0

d e it
= ie e +it

dt t= 0

= ie
e eit +it
(
e it i + i ) t= 0

= ie ( )

e i + i
= i ( + 1) = ( + 1) = E ( X ),
2 2

so that

( )
2 X = E X 2 EX ( ) ( ) 2
(
= + 1 2 = ; )
that is, 2(X ) = .

6.3.2 Continuous Case


1. Let X be N(, 2). Then X(t) = eit( t /2), and, in particular, if X is
22

N(0, 1), then X(t) = et /2. If X is N(, 2), then (X )/, is N(0, 1).
2

Thus

() () ( )
( X ) t = (1 ) X ( ) t = e it X t , and X t = e it ( X ) t . ( ) ()
So it suffices to find the ch.f. of an N(0, 1) r.v. Y, say. Now

( ) 2 dy
() 1 1
y 2 2 ity
e e
2
Y t = ity
e y 2
dy =
2 2
1 (
y it )
2
2
e
2 2
= e t 2
dy = e t 2
.
2

Hence X(t/) = eit/ et /2 and replacing t/ by t, we get, finally:


2

2t 2
()
X t = exp it
2
.
148 6 Characteristic Functions, Moment Generating Functions and Related Theorems

Hence

2t 2
d
dt
()
X t = exp it
2 i t

2
( ) = i , so that E X = . ( )
t =0 t =0

2t 2 2t 2
d2
() ( )
2
X t = exp it i t
2
2 exp it
dt 2 t =0
2 2
t =0

= i2 2 2 = i2 2 + 2 . ( )
Then E(X 2) = 2 + 2 and 2(X) = 2 + 2 2 = 2.
2. Let X be Gamma distributed with parameters and . Then X(t) =
(1 it). In fact,

() 1 1 ( )
x 1 it

( ) 0 ( ) 0
X t = e itx x 1 e x dx = x 1 e dx.

Setting x(1 it) = y, we get

x=
y
1 it
, dx =
dy
1 it
, [
y 0, . )
Hence the above expression becomes

1 1 dy
( ) 0 (1 it ) 1
1 y
y e
1 it

( ) 1
( )

( ) 0
= 1 it y 1e y dy = 1 it .

Therefore

i
d
X t () = = i ,
(1 it )
+1
dt
t =0 t =0

so that E(X ) = , and

d2 (
+1 2 )
X =i 2
(
= i 2 + 1 2 , )
(1 it )
+2
dt 2
t =0 t =0

so that E(X ) = ( + 1) . Thus (X) = ( + 1)2 22 = 2.


2 2 2

For = r/2, = 2, we get the corresponding quantities for 2r, and for
= 1, = 1/, we get the corresponding quantities for the Negative Exponential
distribution. So
6.5 The Moment GeneratingExercises
Function 149

1
it
() ( ) ()
r 2
X t = 1 2it , X t = 1 = ,
it
respectively.
3. Let X be Cauchy distributed with = 0 and = 1. Then X(t) = e|t|. In
fact,

1 cos tx ( )
() 1 1

X t = e itx dx = dx
1+ x 2
1 + x 2
i sin tx
+ dx =
( )
2 cos tx
dx
( )
1+ x 2
0 1 + x2
because
( ) dx = 0,
sin tx

1 + x2

since sin(tx) is an odd function, and cos(tx) is an even function. Further, it can
be shown by complex variables theory that
cos tx ( ) dx = e t

0 1 + x2 2
.

Hence

()
X t = e
t
.
Now
d
dt
X t = e
d t
dt
()
does not exist for t = 0. This is consistent with the fact of nonexistence of E(X ),
as has been seen in Chapter 5.

Exercises
6.3.1 Let X be an r.v. with p.d.f. f given in Exercise 3.2.13 of Chapter 3.
Derive its ch.f. , and calculate EX, E[X(X 1)], 2(X), provided they are
finite.
6.3.2 Let X be an r.v. with p.d.f. f given in Exercise 3.2.14 of Chapter 3.
Derive its ch.f. , and calculate EX, E[X(X 1)], 2(X), provided they are
finite.
6.3.3 Let X be an r.v. with p.d.f. f given by f(x) = e(x) I(,)(x). Find its ch.f.
, and calculate EX, 2(X), provided they are finite.
150 6 Characteristic Functions, Moment Generating Functions and Related Theorems

6.3.4 Let X be an r.v. distributed as Negative Binomial with parameters r


and p.
i) Show that its ch.f., , is given by

pr
()
t = ;
( )
r
1 qe it

ii) By differentiating , show that EX = rq/p and 2(X) = rq/p2;


iii) Find the quantities mentioned in (i) and (ii) for the Geometric
distribution.
6.3.5 Let X be an r.v. distributed as U(, ).
ii) Show that its ch.f., , is given by

e it e it
()
t = ;
(
it )
ii) By differentiating , show that EX = 2+ and 2 ( X ) = ( )2
12
.
6.3.6 Consider the r.v. X with p.d.f. f given in Exercise 3.3.14(ii) of Chapter
3, and by using the ch.f. of X, calculate EX n, n = 1, 2, . . . , provided they are
finite.

6.4 Definitions and Basic TheoremsThe Multidimensional Case


In this section, versions of Theorems 1, 2 and 3 are presented for the case that
the r.v. X is replaced by a k-dimensional r. vector X. Their interpretation,
usefulness and usage is analogous to the ones given in the one-dimensional
case. To this end, let now X = (X1, . . . , Xk) be a random vector. Then the ch.f.
of the r. vector X, or the joint ch.f. of the r.v.s X1, . . . , Xk, denoted by X or
X1, . . . , Xk, is defined as follows:
X , ,
1 Xk (t , , t ) = E[e
1 k
it 1 X 1 +it2 X 2 + + itk X k
], t j ! ,

j = 1, 2, . . . , k. The ch.f. X1, . . . , Xk always exists by an obvious generalization


of Lemmas A, A and B, B. The joint ch.f. X1, . . . , Xk satisfies properties
analogous to properties (i)(vii). That is, one has
THEOREM 1 (Some properties of ch.f.s)
i) X1, . . . , Xk(0, . . . , 0) = 1.
ii) |X1, . . . , Xk(t1, . . . , tk)| 1.
iii) X1, . . . , Xk is uniformly continuous.
6.4 Definitions and Basic TheoremsThe
6.5 The Moment
Multidimensional
Generating Function
Case 151

iv) X1+d1, . . . , Xk+dk(t1, . . . , tk) = eit1d1+ +itkdkX1, . . . , Xk (t1, . . . , tk).


v) c1X1, . . . , ckXk(t1, . . . , tk) = X1, . . . , Xk (c1t1, . . . , cktk).
it1d1 + + itkdk
vi) c1X1+ d1, . . . , ckXk + dk(t1, . . . , tk) = e X1, . . . , Xk(c1t1, . . . , cktk).
vii) If the absolute (n1, . . . , nk)-joint moment, as well as all lower order joint
moments of X1, . . . , Xk are finite, then

n + +n
(t , , t ) = i ( )
1 k k
j=1 n j
X , , Xk 1 k E X 1n X kn ,
1 k

t tn t tn
1 k 1

1 k t 1 = =tk = 0

and, in particular,

n
t nj
X , , 1 Xk (t , , t )
1 k ( )
= i n E X nj , j = 1, 2, , k.
t 1 = =tk = 0

viii) If in the X1, . . . , Xk(t1, . . . , tk) we set tj1 = = tjn = 0, then the resulting
expression is the joint ch.f. of the r.v.s Xi1, . . . , Xim, where the js and the
is are different and m + n = k.
Multidimensional versions of Theorem 2 and Theorem 3 also hold true.
We give their formulations below.

THEOREM 2 (Inversion formula) Let X = (X1, . . . , Xk) be an r. vector with p.d.f. f and ch.f.
. Then
k
1
( )
T T
ii) f X , ,
1 Xk x1 , , xk = lim
T 2T T
e it x it x
T
1 1 k k

X , ,
1 Xk (t , , t )dt
1 k 1 dt k ,

if X is of the discrete type, and


k
1 k 1 e it j h
(x , , x )
T T
ii) f X , ,
1 Xk 1 k = lim lim
h 0 T 2 T

T
j =1 it j h

e it1x1 itk xk X 1 , , Xk (t , , t )dt


1 k 1 dt k ,

if X is of the continuous type, with the analog of (ii) holding if the integral
of |X1, . . . , Xk(t1, . . . , tk)| is finite.
THEOREM 3 (Uniqueness Theorem) There is a one-to-one correspondence between the
ch.f. and the p.d.f. of an r. vector.
PROOFS The justification of Theorem 1 is entirely analogous to that given
for Theorem 1, and so is the proof of Theorem 2. As for Theorem 3, the fact
that the p.d.f. of X determines its ch.f. follows from the definition of the ch.f.
That the ch.f. of X determines its p.d.f. follows from Theorem 2. "
152 6 Characteristic Functions, Moment Generating Functions and Related Theorems

6.4.1 The Ch.f. of the Multinomial Distribution


Let X = (X1, . . . , Xk) be Multinomially distributed; that is,

(
P X 1 = x1 , , X k = xk = ) n!
x1! xk !
p1x pkx . 1 k

Then

(t , , t ) = ( p e ).
n
it 1
X , ,
1 Xk 1 k 1 + + pk e it k

In fact,

X , ,
1 Xk (t , , t ) =
1 k
x 1 , , xk
e it x + +it x
1 1 k k
n!
x1! xk !
p1x pkx 1 k

n!
( ) ( )
x1 xk
= x1! xk !
p1e it 1
pk e it k

x 1 , , xk

( ).
n
= p1e it + + pk e it
1 k

Hence

k
t1 t k
X , ,1 Xk (t , , t )
1 k
t 1 = = tk = 0

( ) ( )
= n n 1 n k + 1 i p1 pk p1e it1 + k
(
) ( ) ( )
n k
+ pk e itk = i k n n 1 n k + 1 p1 p2 pk .
t 1 = = tk = 0

Hence
( ) (
E X1 X k = n n 1 n k + 1 p1 p2 pk .) ( )
Finally, the ch.f. of a (measurable) function g(X) of the r. vector X =
(X1, . . . , Xk) is defined by:
e itg ( x ) f x , x = x ,
()
1 , xk ( )
()
g(X) t = E e

itg ( X )
= x

( )
itg ( x , , x )


e f x1 , , xk dx1 , , dxk .
1 k

Exercise
6.4.1 (CramrWold) Consider the r.v.s Xj, j = 1, . . . , k and for cj !,
j = 1, . . . , k, set
6.5 The Moment Generating Function 153

k
Yc = c j X j .
j =1

Then
ii) Show that Yc(t) = X1, . . . , Xk(c1t, . . . , ckt), t !, and X1, . . . , Xk(c1, . . . , ck)
= Yc(1);
ii) Conclude that the distribution of the Xs determines the distribution of Yc
for every cj !, j = 1, . . . , k. Conversely, the distribution of the Xs is
determined by the distribution of Yc for every cj !, j = 1, . . . , k.

6.5 The Moment Generating Function and Factorial Moment


Generating Function of a Random Variable
The ch.f. of an r.v. or an r. vector is a function defined on the entire real line
and taking values in the complex plane. Those readers who are not well versed
in matters related to complex-valued functions may feel uncomfortable in
dealing with ch.f.s. There is a partial remedy to this potential problem, and
that is to replace a ch.f. by an entity which is called moment generating
function. However, there is a price to be paid for this: namely, a moment
generating function may exist (in the sense of being finite) only for t = 0. There
are cases where it exists for ts lying in a proper subset of ! (containing 0), and
yet other cases, where the moment generating function exists for all real t. All
three cases will be illustrated by examples below.
First, consider the case of an r.v. X. Then the moment generating function
(m.g.f.) MX (or just M when no confusion is possible) of a random variable X,
which is also called the Laplace transform of f, is defined by MX(t) = E(etX),
t !, if this expectation exists. For t = 0, MX(0) always exists and equals
1. However, it may fail to exist for t 0. If MX(t) exists, then formally
X(t) = MX(it) and therefore the m.g.f. satisfies most of the properties analo-
gous to properties (i)(vii) cited above in connection with the ch.f., under
suitable conditions. In particular, property (vii) in Theorem 1 yields
M X (t ) t =0 = E ( X n ) , provided Lemma D applies. In fact,
n
d
n
dt

dn
dn
dt n
MX t () =
dn
dt n
(
Ee tX ) = E n e tX
dt t =0
t =0 t =0

(
= E X n e tX ) t =0
( )
= E Xn .

This is the property from which the m.g.f. derives its name.
154 6 Characteristic Functions, Moment Generating Functions and Related Theorems

Here are some examples of m.g.f.s. It is instructive to derive them in order


to see how conditions are imposed on t in order for the m.g.f. to be finite. It so
happens that part (vii) of Theorem 1, as it would be formulated for an m.g.f.,
is applicable in all these examples, although no justification will be supplied.

6.5.1 The M.G.F.s of Some R.V.s


1. If X B(n, p), then MX(t) = (pet + q)n, t !. Indeed,

n n
( ) ( )
n n

()
x n
M X t = e tx p x q n x = pe t q n x = pe t + q ,
x= 0 x x = 0 x

which, clearly, is finite for all t !.


Then

d
() d
( ) ( ) ( )
n n 1
MX t = pe t + q = n pe t + q pe t = np = E X ,
dt t =0
dt t =0 t =0

and

d2
() d
( )
n 1
MX t = np pe t + q et
dt 2 t =0
dt t =0

( )( ) ( )
n 2 n 1
= np n 1 pe t + q pe t e t + pe t + q et
t =0
( )
= n n 1 p 2 + np = n 2 p 2 np 2 + np = E X 2 , ( )
so that 2(X) = n2p2 np2 + np n2p2 = np(1 p) = npq.
2. If X P(), then MX(t) = ee , t !. In fact,
t

( )
x
e
t

x
()
M X t = e tx e =e
t t

= e e e = e e .
x= 0 x! x= 0 x!
Then

d
() d e
( )
t t
MX t = e = e t e e ==E X ,
dt t =0
dt t =0
t =0

and

d2
dt 2
()
MX t
t =0
=
d
dt
( t
e t e e ) t =0
( t
= e t e e + e t e e e t
t

) t =0

( ) ( )
= 1 + = E X 2 , so that 2 X = + 2 2 = . ( )
6.5 The Moment Generating Function 155

2t 2
3. If X N(, 2), then M X (t ) = e
t +
2
, t !, and, in particular, if X N(0,
1), then MX(t) = et /2, t !. By the property for m.g.f. analogous to property (vi)
2

in Theorem 1,

t t
()
M X t = e t M X , so that M X = e t M X t

()

for all t !. Therefore


t t t 2
t 2
2
+
MX = e et =e 2
.

2t 2
Replacing t by t, we get, finally, M X (t ) = e
t +
2
. Then

2t 2 2t 2
d
() d t +
( ) ( )
t +
MX t = e 2
= + 2t e 2
==E X ,
dt t= 0
dt
t=0 t= 0

and

d
2t 2
d2
() ( )
t +
MX t = + 2t e 2

dt 2 dt
t=0 t= 0

t 2 2
t 2 2

( )
t + 2 t +
= 2 e 2
+ + 2t e 2
= 2 + 2

t=0

( )
= E X 2 , so that 2 X = 2 + 2 2 = 2 . ( )
4. If X is distributed as Gamma with parameters and , then MX(t) =
(1 t), t < 1/. Indeed,

() 1 1 ( )
x 1 t

( ) ( )
MX t = e tx x 1 e x dx = x 1 e dx.
0
0

Then by setting x(1 t) = y, so that x = y


1 t
, dx = dy
1 t
, and y [0, ), the above
expression is equal to

1 1 1
( ) 0
y 1 e y dy = ,
(1 t ) (1 t )

provided 1 t > 0, or equivalently, t < 1/. Then

d
() d
( ) ( )

MX t = 1 t = = E X ,
dt t =0
dt t =0
156 6 Characteristic Functions, Moment Generating Functions and Related Theorems

and

d2
() d
( ) ( ) ( )
1 2
MX t = 1 t = + 1 2 1 t
dt 2 t =0
dt t =0 t =0

( ) (
= + 1 = EX , so that X = .
2 2
) 2
( ) 2

In particular, for = 2r and = 2, we get the m.g.f. of the 2r, and its mean and
variance; namely,

() ( 1
) ( ) ( )
r 2
MX t = 1 2t , E X = r , 2 X = 2r .
, t<
2
For = 1 and = 1 , we obtain the m.g.f. of the Negative Exponential
distribution, and its mean and variance; namely

MX t =() t
1
, t < , EX = , 2 X = 2 .

1
( )
5. Let X have the Cauchy distribution with parameters and , and
without loss of generality, let = 0, = 1. Then the MX(t) exists only for t = 0.
In fact.

() ( )
1 1

M X t = E e tX = e tx dx
1 + x2
1
> e tx
0
1
1+ x 2
1
dx > tx
0
1
1 + x2
dx ( )
if t > 0, since ez > z, for z > 0, and this equals
t
2
2 x dx
1+ x

20
=
t du
2 1 u
=
t
lim log u .
2 x ( )
Thus for t > 0, MX(t) obviously is equal to . If t < 0, by using the limits , 0
in the integral, we again reach the conclusion that MX(t) = (see Exercise
6.5.9).
REMARK 4 The examples just discussed exhibit all three cases regarding the
existence or nonexistence of an m.g.f. In Examples 1 and 3, the m.g.f.s exist
for all t !; in Examples 2 and 4, the m.g.f.s exist for proper subsets of !; and
in Example 5, the m.g.f. exists only for t = 0.
For an r.v. X, we also define what is known as its factorial moment
generating function. More precisely, the factorial m.g.f. X (or just when no
confusion is possible) of an r.v. X is defined by:
() ( )
X t = E t X , t ! , if E t X exists. ( )
This function is sometimes referred to as the Mellin or MellinStieltjes trans-
form of f. Clearly, X(t) = MX(log t) for t > 0.
Formally, the nth factorial moment of an r.v. X is taken from its factorial
m.g.f. by differentiation as follows:
6.5 The Moment Generating Function 157

dn
dt n
X t ()
t =1
[ ( ) (
= E X X 1 X n+1 . )]
In fact,

n X
dn
dt n
()
X t =
dn
dt n
E t X
=( )
E n t = E X X 1 X n+1 t
t
[ (
X n
, ) ( ) ]
provided Lemma D applies, so that the interchange of the order of differen-
tiation and expectation is valid. Hence

dn
dt n
X t ()
t =1
[ ( ) (
= E X X 1 X n+1 . )] (9)

REMARK 5 The factorial m.g.f. derives its name from the property just estab-
lished. As has already been seen in the first two examples in Section 2 of
Chapter 5, factorial moments are especially valuable in calculating the vari-
ance of discrete r.v.s. Indeed, since

( ) ( ) ( ) 2
2 X = E X 2 EX , and E X 2 = E X X 1 + E X , ( ) [ ( )] ( )
we get

( ) [ (
2 X = E X X 1 + E X EX ; )] ( ) ( )
2

that is, an expression of the variance of X in terms of derivatives of its factorial


m.g.f. up to order two.
Below we derive some factorial m.g.f.s. Property (9) (for n = 2) is valid in
all these examples, although no specific justification will be provided.

6.5.2 The Factorial M.G.F.s of some R.V.s


1. If X B(n, p), then X(t) = (pt + q)n, t !. In fact,
n
n n
n
() ( ) ( )
x n
X t = t x p x q n x = pt q n x = pt + q .
x= 0 x x= 0 x
Then

d2
() ( ) ( ) ( )
n 2
X t = n n 1 p 2 pt + q = n n 1 p2 ,
dt 2 t =1

so that 2(X) = n(n 1)p2 + np n2p2 = npq.


2. If X P(), then X(t) = et , t !. In fact,

( )
x
t

x
()
X t = t e
x =0
x

x!
= e
x = 0 x!
= e e t = e t , t ! .
158 6 Characteristic Functions, Moment Generating Functions and Related Theorems

Hence
d2
dt 2
X t () = 2 e t
t =1
= 2 , so that 2 X = 2 + 2 = . ( )
t =1

The m.g.f. of an r. vector X or the joint m.g.f. of the r.v.s X1, . . . , Xk,
denoted by MX or MX1, . . . , Xk, is defined by:

1 k
( ) (
M X , , X t1 , , t k = E e t X + t X , t j ! , j = 1, 2, , k,
1 1 k k
)
for those tjs in ! for which this expectation exists. If MX1, . . . , Xk(t1, . . . , tk) exists,
then formally X1, . . . , Xk(t1, . . . , tk) = MX1, . . . , Xk(it1, . . . , itk) and properties analo-
gous to (i)(vii), (viii) in Theorem 1 hold true under suitable conditions. In
particular,

n + +n
(t , , t ) ( )
1 k

MX , , Xk 1 k = E X 1n 1 X knk , (10)
t1n 1 t knk 1

t 1 = = tk = 0

where n1, . . . , nk are non-negative integers.


Below, we present two examples of m.g.f.s of r. vectors.

6.5.3 The M.G.F.s of Some R. Vectors


1. If the r.v.s X1, . . . , Xk have jointly the Multinomial distribution with
parameters n and p1, . . . , pk, then

(t , , t ) = ( p e ),
n
t1
MX , ,
1 Xk 1 k 1 + + pk e t k
t j ! , j = 1, , k.

In fact,

MX1 , , Xk (t , , t ) = Ee
1 k
t 1 X 1 + +tk X k
= e 1
t X 1 + +tk X k n!
x1! xk !
p1x1 pkxk

n!
( ) ( )
x1 xk
= p1e t1 pk e tk
x1! xk !

( ),
n
= p1e t1 + + pk e tk

where the summation is over all integers x1, . . . , xk 0 with x1 + + xk = n.


Clearly, the above derivations hold true for all tj !, j = 1, . . . , k.
2. If the r.v.s X1 and X2 have the Bivariate Normal distribution with
parameters 1, 2, 21, 22 and , then their joint m.g.f. is

MX 1 ,X 2
(t , t ) = exp t
1 2


1 1 + 2t 2 +
2
(
1 2 2
1 t1 + 2 1 2 t1t 2 + 22 t 22 ), t , t
1 2 ! . (11)

An analytical derivation of this formula is possible, but we prefer to use


the matrix approach, which is more elegant and compact. Recall that the joint
p.d.f. of X1 and X2 is given by
6.5 The Moment Generating Function 159

(
f x1 , x 2 )
2
1 x1 1 x1 1 x 2 2 x 2 2
2
1
= exp 2 + .
2 1 2 1 2 2 1 1 2 2


Set x = (x1 x2), = (1 2), and

12
12 2 .
=
1 2 2

Then the determinant of , | | = 21 22(1 2), and the inverse, 1, is


|, is |

1 2 1 2
2
1 = .
1 2 12

Therefore

22 1 2 x1 1
( )
x 1 x = ( ) 1

(
x1 1 x 2 2
1 2
)
12 x 2 2
1
(
2 x
) ( )( ) ( )
2 1 2 x1 1 x 2 2 + 12 x 2 2
2 2
=
1
1
2 2
2 ( 2
)
2 1 1

1 x 2 x1 1 x 2 2 x 2 2
2

= 1 1
2 +
.
1 2 1 1 2 2

Therefore the p.d.f. is written as follows in matrix notation:

1
()
f x =
1
( )
exp x 1 x ( ) .
2
1 2
2

In this form, is the mean vector of X = (X1 X2), and is the covariance matrix
of X.
Next, for t = (t1 t2), we have

()
M X t = Ee t X = 2 exp t x f x dx
!
( )()

! exp t x 2 ( x ) ( x )dx.
1 1 1
= 1 2 2
(11)
2

The exponent may be written as follows:


160 6 Characteristic Functions, Moment Generating Functions and Related Theorems



1
2
1
2
[ ( 1
)
t + t t 2 t + t t 2 t x + x x . ( )] (13)

Focus on the quantity in the bracket, carry out the multiplication, and observe
1) = 1, xt = tx, t = t, and x 1 =
that = , ( 1x, to obtain

( ) ( ) [ ( ) ( (
2 t + t t 2t x + x 1 x = x + t 1 x + t . ))] (14)

By means of (13) and (14), the m.g.f. in (12) becomes

MX t()
1
= exp t + t t
2
1
2 1 2 ! 2
1
( ( )) ( (
exp x + t 1 x + t dx.
2
))
However, the second factor above is equal to 1, since it is the integral of a
Bivariate Normal distribution with mean vector + t and covariance matrix
. Thus


()
1
M X t = exp t + t t .
2
(15)

Observing that

12 1 2 t1
( )
t t = t1 t 2 = 1 t1 + 2 1 2 t1t 2 + 2 t 2 ,
1 2 2 t 2
2
2 2 2 2

it follows that the m.g.f. is, indeed, given by (11).

Exercises
6.5.1 Derive the m.g.f. of the r.v. X which denotes the number of spots that
turn up when a balanced die is rolled.

6.5.2 Let X be an r.v. with p.d.f. f given in Exercise 3.2.13 of Chapter 3.


Derive its m.g.f. and factorial m.g.f., M(t) and (t), respectively, for those ts
for which they exist. Then calculate EX, E[X(X 1)] and 2(X), provided they
are finite.

6.5.3 Let X be an r.v. with p.d.f. f given in Exercise 3.2.14 of Chapter 3.


Derive its m.g.f. and factorial m.g.f., M(t) and (t), respectively, for those ts
for which they exist. Then calculate EX, E[X(X 1)] and 2(X), provided they
are finite.
6.5 The Moment GeneratingExercises
Function 161

6.5.4 Let X be an r.v. with p.d.f. f given by f(x) = e(x )I(,)(x). Find its
m.g.f. M(t) for those ts for which it exists. Then calculate EX and 2(X),
provided they are finite.
6.5.5 Let X be an r.v. distributed as B(n, p). Use its factorial m.g.f. in order
to calculate its kth factorial moment. Compare with Exercise 5.2.1 in Chapter
5.
6.5.6 Let X be an r.v. distributed as P(). Use its factorial m.g.f. in order to
calculate its kth factorial moment. Compare with Exercise 5.2.4 in Chapter 5.
6.5.7 Let X be an r.v. distributed as Negative Binomial with parameters
r and p.
i) Show that its m.g.f and factorial m.g.f., M(t) and (t), respectively, are
given by

pr pr
()
MX t = , t < log q, X t = () , t <
1
;
(1 qe ) (1 qt ) q
r r
t

ii) By differentiation, show that EX = rq/p and 2(X) = rq/p2;


iii) Find the quantities mentioned in parts (i) and (ii) for the Geometric
distribution.
6.5.8 Let X be an r.v. distributed as U(, ).
ii) Show that its m.g.f., M, is given by

e t e t
Mt = () ;
(
t )
ii) By differentiation, show that EX = +2 and 2(X) = ( ) .
2

12

6.5.9 Refer to Example 3 in the Continuous case and show that MX(t) = for
t < 0 as asserted there.
2
6.5.10 Let X be an r.v. with m.g.f. M given by M(t) = e t + t , t ! ( !,
> 0). Find the ch.f. of X and identify its p.d.f. Also use the ch.f. of X in order
to calculate EX4.
6.5.11 For an r.v. X, define the function by (t) = E(1 + t)X for those ts for
which E(1 + t)X is finite. Then, if the nth factorial moment of X is finite, show
that

(d n
) ()
dt n t
t =0
[ ( )
= E X X 1 X n+1 .( )]
6.5.12 Refer to the previous exercise and let X be P(). Derive (t) and use
it in order to show that the nth factorial moment of X is n.
162 6 Characteristic Functions, Moment Generating Functions and Related Theorems

6.5.13 Let X be an r.v. with m.g.f. M and set K(t) = log M(t) for those ts for
which M(t) exists. Furthermore, suppose that EX = and 2(X) = 2 are both
finite. Then show that
d2
d
dt
Kt() = and
dt 2
Kt () = 2.
t =0 t =0

(The function K just defined is called the cumulant generating function of X.)

6.5.14 Let X be an r.v. such that EX n is finite for all n = 1, 2, . . . . Use the
expansion

xn
ex =
n = 0 n!

in order to show that, under appropriate conditions, one has that the m.g.f. of
X is given by

( ) tn! .
n

()
M t = EX n
n =0

6.5.15 If X is an r.v. such that EX n = n!, then use the previous exercise in
order to find the m.g.f. M(t) of X for those ts for which it exists. Also find the
ch.f. of X and from this, deduce the distribution of X.
6.5.16 Let X be an r.v. such that

EX 2k =
(2k)! , EX 2k +1 = 0,
2 k k!
k = 0, 1, . . . . Find the m.g.f. of X and also its ch.f. Then deduce the distribution
of X. (Use Exercise 6.5.14)
6.5.17 Let X1, X2 be two r.v.s with m.f.g. given by
2

( )
1
3
( 1
) (
)
M t1 , t 2 = e t1 + t2 + 1 + e t1 + e t2 , t1 , t 2 ! .
6
Calculate EX1, 2(X1) and Cov(X1, X2), provided they are finite.
6.5.18 Refer to Exercise 4.2.5. in Chapter 4 and find the joint m.g.f.
M(t1, t2, t3) of the r.v.s X1, X2, X3 for those t1, t2, t3 for which it exists. Also find
their joint ch.f. and use it in order to calculate E(X1X2X3), provided the
assumptions of Theorem 1 (vii) are met.

6.5.19 Refer to the previous exercise and derive the m.g.f. M(t) of the r.v.
g(X1, X2, X3) = X1 + X2 + X3 for those ts for which it exists. From this, deduce
the distribution of g.
6.5 The Moment GeneratingExercises
Function 163

6.5.20 Let X1, X2 be two r.v.s with m.g.f. M and set K(t1, t2) = log M(t1, t2) for
those t1, t2 for which M(t1, t2) exists. Furthermore, suppose that expectations,
variances, and covariances of these r.v.s are all finite. Then show that for
j = 1, 2,

2
t j
(
K t1 , t 2 ) = EX j ,
t j2
(
K t1 , t 2 ) ( )
= 2 Xj ,
t 1 =t2 = 0 t 1 =t2 = 0

2
t1t 2
K t1 , t 2 ( ) (
= Cov X 1 , X 2 . )
t 1 =t2 = 0

6.5.21 Suppose the r.v.s X1, . . . , Xk have the Multinomial distribution with
parameters n and p1, . . . , pk, and let i, j, be arbitrary but fixed, 1 i < j k.
Consider the r.v.s Xi, Xj, and set X = n Xi Xj, so that these r.v.s have the
Multinomial distribution with parameters n and pi, pj, p, where p = 1 pi pj.
ii) Write out the joint m.g.f. of Xi, Xj, X, and by differentiation, determine the
E(XiXj);
ii) Calculate the covariance of Xi, Xj, Cov(Xi, Xj), and show that it is negative.
6.5.22 If the r.v.s X1 and X2 have the Bivariate Normal distribution with
parameters 1, 2, 21, 22 and , show that Cov(X1, X2) 0 if 0, and
Cov(X1, X2) < 0 if < 0. Note: Two r.v.s X1, X2 for which Fx ,x (X1, X2) 1 2
Fx (X1)Fx (X2) 0, for all X1, X2 in !, or Fx ,x (X1, X2) Fx (X1)Fx (X2) 0, for
1 2 1 2 1 2
all X1, X2 in !, are said to be positively quadrant dependent or negatively
quadrant dependent, respectively. In particular, if X1 and X2 have the Bivariate
Normal distribution, it can be seen that they are positively quadrant depend-
ent or negatively quadrant dependent according to whether 0 or < 0.
6.5.23 Verify the validity of relation (13).
6.5.24
ii) If the r.v.s X1 and X2 have the Bivariate Normal distribution with param-
eters 1, 2, 21, 22 and , use their joint m.g.f. given by (11) and property
(10) in order to determine E(X1X2);
ii) Show that is, indeed, the correlation coefficient of X1 and X2.
6.5.25 Both parts of Exercise 6.4.1 hold true if the ch.f.s involved are re-
placed by m.g.f.s, provided, of course, that these m.g.f.s exist.

ii) Use Exercise 6.4.1 for k = 2 and formulated in terms of m.g.f.s in order to
show that the r.v.s X1 and X2 have a Bivariate Normal distribution if and
only if for every c1, c2 !, Yc = c1X1 + c2X2 is normally distributed;
ii) In either case, show that c1X1 + c2X2 + c3 is also normally distributed for any
c3 !.
164 7 Stochastic Independence with Some Applications

Chapter 7

Stochastic Independence with Some


Applications

7.1 Stochastic Independence: Criteria of Independence


Let S be a sample space, consider a class of events associated with this space,
and let P be a probability function defined on the class of events. In Chapter
2 (Section 2.3), the concept of independence of events was defined and was
heavily used there, as well as in subsequent chapters. Independence carries
over to r.v.s also, and is the most basic assumption made in this book. Inde-
pendence of r.v.s, in essence, reduces to that of events, as will be seen below.
In this section, the not-so-rigorous definition of independence of r.v.s is pre-
sented, and two criteria of independence are also discussed. A third criterion
of independence, and several applications, based primarily on independence,
are discussed in subsequent sections. A rigorous treatment of some results is
presented in Section 7.4.
DEFINITION 1 The r.v.s Xj, j = 1, . . . , k are said to be independent if, for sets Bj !, j = 1, . . . ,
k, it holds
k

( ) (
P X j Bj , j = 1, , k = P X j Bj .
j =1
)
The r.v.s Xj, j = 1, 2, . . . are said to be independent if every finite subcollection
of them is a collection of independent r.v.s. Non-independent r.v.s are said to
be dependent. (See also Definition 3 in Section 7.4, and the comment following
it.)
REMARK 1 (i) The sets Bj, j = 1, . . . , k may not be chosen entirely arbitrar-
ily, but there is plenty of leeway in their choice. For example, taking Bj = (,
xj], xj !, j = 1, . . . , k would be sufficient. (See Lemma 3 in Section 7.4.)
(ii) Definition 1 (as well as Definition 3 in Section 7.4) also applies to m-
dimensional r. vectors when ! (and B in Definition 3) is replaced by ! m (Bm).

164
7.1 Stochastic Independence: Criteria of Independence 165

THEOREM 1 (Factorization Theorem) The r.v.s Xj, j = 1, . . . , k are independent if and only
if any one of the following two (equivalent) conditions holds:
k

1 k
( )
i) FX , , X x1 , , xk = FX x j , for all
j =1
j
( ) x j ! , j = 1, , k.

1 k
( )
ii) f X , , X x1 , , xk = f X x j , for all
j =1
j
( ) x j ! , j = 1, , k.
PROOF
ii) If Xj, j = 1, , k are independent, then
k

( )
P X j Bj , j = 1, , k = P X j Bj , Bj ! , j = 1, , k.
j =1
( )
In particular, this is true for Bj = (, xj], xj !, j = 1, . . . , k which gives
k

1 k
(
FX , , X x1 , , xk = FX x j . ) j =1
j
( )
The proof of the converse is a deep probability result, and will, of course,
be omitted. Some relevant comments will be made in Section 7.4, Lemma 3.
ii) For the discrete case, we set Bj = {xj}, where xj is in the range of Xj, j = 1, . . . ,
k. Then if Xj, j = 1, . . . , k are independent, we get
k

(
P X 1 = x1 , , X k = xk = P X j = x j , ) j =1
( )
or
k

1 k
(
f X , , X x1 , , xk = f X x j . ) j =1
j
( )
Let now
k

1 k
(
f X , , X x1 , , xk = f X x j . ) j =1
j
( )
Then for any sets Bj = (, yj], yj !, j = 1, . . . , k, we get

B1

B
1 k
(
f X , , X x1 , , xk = ) B1

B
1
( )
f X x1 f X xk k
( )
k k

k
= f X x j ,
j =1
j
( )
B j

or
k

1 k
(
FX , , X y1 , , yk = FX y j . ) j =1
j
( )
Therefore Xj, j = 1, . . . , k are independent by (i). For the continuous case,
we have: Let
k

1 k
(
f X , , X x1 , , xk = f X x j ) j =1
j
( )
and let
166 7 Stochastic Independence with Some Applications

( ]
C j = , y j , y j ! . j = 1, , k.
Then integrating both sides of this last relationship over the set C1
Ck, we get
k

1 k
(
FX , , X y1 , , yk = FX y j , ) j =1
j
( )
so that Xj, j = 1, . . . , k are independent by (i). Next, assume that
k

1 k
(
FX , , X x1 , , xk = FX x j ) j =1
j
( )
(that is, the Xjs are independent). Then differentiating both sides, we get
k

1 k
( )
f X , , X x1 , , xk = f X x j .
j =1
j
( )
REMARK 2 It is noted that this step also is justifiable (by means of calculus)
for the continuity points of the p.d.f. only.
Consider independent r.v.s and suppose that gj is a function of the jth r.v.
alone. Then it seems intuitively clear that the r.v.s gj(Xj), j = 1, . . . , k ought to
be independent. This is, actually, true and is the content of the following
LEMMA 1 For j = 1, . . . , k, let the r.v.s Xj be independent and consider (measurable)
functions gj : ! !, so that gj(Xj), j = 1, . . . , k are r.v.s. Then the r.v.s gj(Xj),
j = 1, . . . , k are also independent. The same conclusion holds if the r.v.s are
replaced by m-dimensional r. vectors, and the functions gj, j = 1, . . . , k are
defined on ! m into !. (That is, functions of independent r.v.s (r. vectors)
are independent r.v.s.)
PROOF See Section 7.4.
Independence of r.v.s also has the following consequence stated as a
lemma. Both this lemma, as well as Lemma 1, are needed in the proof of
Theorem 1 below.
LEMMA 2 Consider the r.v.s Xj, j = 1, . . . , k and let gj : ! ! be (measurable) functions,
so that gj(Xj), j = 1, . . . , k are r.v.s. Then, if the r.v.s Xj, j = 1, . . . , k are
independent, we have
k k
( )
E g j X j = E g j X j ,
j = 1 j = 1
[ ( )]
provided the expectations considered exist. The same conclusion holds if the
gjs are complex-valued.
PROOF See Section 7.2.
REMARK 3 The converse of the above statement need not be true as will be
seen later by examples.
THEOREM 1 (Factorization Theorem) The r.v.s Xj, j = 1, . . . , k are independent if and only if:
k

1 k
( ) j =1
j
( )
X , , X t1 , , t k = X t j , for all t j ! , j = 1, , k.
7.1 Stochastic Independence: Criteria of Independence 167

PROOF If X1, j = 1, . . . , k are independent, then by Theorem 1(ii),


k

1 k
(
f X , , X x1 , , xk = f X x j . ) j =1
j
( )
Hence
k k it X k

(
)
X , , X x1 , , xk = E exp i t j X j = E e = Ee

it X j j j j


1 k
j =1 j =1 j =1

by Lemmas 1 and 2, and this is kj=1X (tj). Let us assume now that j
k

1 k
(
X , , X t1 , , t k = X t j . ) j =1
j
( )
For the discrete case, we have (see Theorem 2(i) in Chapter 6)

f X x j = lim j
( )
1 T it x
T 2T T
e X t j dt j , j = 1, , k, j j

j
( )
and for the multidimensional case, we have (see Theorem 2(i) in Chapter 6)

1
k
k
f X 1 , , X k ( x1 , , x k ) = lim
T T

exp i t j x j
T 2T T T
j =1
X 1 , , X k (t1 , , t k )dt1 dt k
1
k
k k
exp i t j x j X j t j ( dt1 dt k ) ( )
T T
= lim
T 2T
T T
j =1 j =1
1 T it j x j k
( )
k
= lim e X j t j dt j = f X x .
j =1
T 2T T
j =1 j ( j )

That is, Xj, j = 1, . . . , k are independent by Theorem 1(ii). For the continuous
case, we have
it j h
1 T 1 e
( )
h 0 T 2 T
f X x j = lim lim
j
it j h
it h
e X t j dt j , j = 1, , k, j

j
( )
and for the multidimensional case, we have (see Theorem 2(ii) in Chapter 6)
k
1 k 1 e it j h
( )
T T
T
it j x j
f X 1 , , X k x1 , , xk = lim lim e
h 0 T 2 T it j h
j =1
(
X 1 , , X k t1 , , t k dt1 dt k )
k
1 1 e it j h it j x j
k
( )
T T
= lim lim
h 0 T 2 T e Xj t j
it j h
T
j =1
dt1 dt k
k it h
1 e j it j x j
1
( )
T
= lim lim T it j h e X j t j dt j
h 0 T 2
j =1

k
= fX j ( xj ) ,
j =1
168 7 Stochastic Independence with Some Applications

which again establishes independence of Xj, j = 1, . . . , k by Theorem


1(ii).
REMARK 4 A version of this theorem involving m.g.f.s can be formulated, if
the m.g.f.s exist.
COROLLARY Let X1, X2 have the Bivariate Normal distribution. Then X1, X2 are indepen-
dent if and only if they are uncorrelated.
PROOF We have seen that (see Bivariate Normal in Section 3.3 of Chapter 3)

fX 1, X2 (x , x ) =
1 2
1
e q 2 ,
2 1 2 1 2
where

1 x 2 x1 2 x2 2 x2 2
2

q= 1 1
2
,
1 2 1 1 2 2

and
x 2
( ) x
( )
2

( )
fX 1 x1 =
1
exp

1

2 2
1 , f x =
X2 2
1
exp 2
2
( )
2
2 .

2 1 1
2 2 2

Thus, if X1, X2 are uncorrelated, so that = 0, then
( ) ( )
fX 1 , X 2 x1 , x 2 = fX 1 x1 fX 2 x 2 , ( )
that is, X1, X2 are independent. The converse is always true by Corollary 1 in
Section 7.2.

Exercises
7.1.1 Let Xj, j = 1, . . . , n be i.i.d. r.v.s with p.d.f. f and d.f. F. Set

(
X (1) = min X 1 , , X n , ) (
X n = max X 1 , , X n ; )
that is,

() [ ()
X (1) s = min X 1 s , , X n s , ( )] () [ () ( )]
X ( n ) s = max X 1 s , , X n s .
Then express the d.f. and p.d.f. of X(1), X(n) in terms of f and F.
7.1.2 Let the r.v.s X1, X2 have p.d.f. f given by f(x1, x2) = I(0,1) (0,1)(x1, x2).
ii) Show that X1, X2 are independent and identify their common distribution;
ii) Find the following probabilities: P(X1 + X2 < 13 ), P( X12 + X 22 < 14 ),
P(X1X2 > 12 ).
7.1.3 Let X1, X2 be two r.v.s with p.d.f. f given by f(x1, x2) = g(x1)h(x2).
Exercises
7.1 Stochastic Independence: Criteria of Independence 169

ii) Derive the p.d.f. of X1 and X2 and show that X1, X2 are independent;
ii) Calculate the probability P(X1 > X2) if g = h and h is of the continuous
type.
7.1.4 Let X1, X2, X3 be r.v.s with p.d.f. f given by f(x1, x2, x3) = 8x1x2x3 IA(x1,
x2, x3), where A = (0, 1) (0, 1) (0, 1).
ii) Show that these r.v.s are independent;
ii) Calculate the probability P(X1 < X2 < X3).
7.1.5 Let X1, X2 be two r.v.s with p.d.f f given by f(x1, x2) = cIA(x1, x2), where
A = {(x1, x2) !2; x12 + x 22 9}.
ii) Determine the constant c;
ii) Show that X1, X2 are dependent.
7.1.6 Let the r.v.s X1, X2, X3 be jointly distributed with p.d.f. f given by

( )1
(
f x1 , x 2 , x3 = I A x1 , x 2 , x3 ,
4
)
where

{( )( )( )(
A = 1, 0, 0 , 0, 1, 0 , 0, 0, 1 , 1, 1, 1 . )}
Then show that
ii) Xi, Xj, i j, are independent;
ii) X1, X2, X3 are dependent.
7.1.7 Refer to Exercise 4.2.5 in Chapter 4 and show that the r.v.s X1, X2, X3
are independent. Utilize this result in order to find the p.d.f. of X1 + X2 and X1
+ X2 + X3 .
7.1.8 Let Xj, j = 1, . . . , n be i.i.d. r.v.s with p.d.f. f and let B be a (Borel) set
in !.
iii) In terms of f, express the probability that at least k of the Xs lie in B for
some fixed k with 1 k n;
iii) Simplify this expression if f is the Negative Exponential p.d.f. with param-
eter and B = (1/, );
iii) Find a numerical answer for n = 10, k = 5, = 12 .
7.1.9 Let X1, X2 be two independent r.v.s and let g: ! ! be measurable.
Let also Eg(X2) be finite. Then show that E[g(X2) | X1 = x1] = Eg(X2).
7.1.10 If Xj, j = 1, . . . , n are i.i.d. r.v.s with ch.f. and sample mean X ,
express the ch.f. of X in terms of .
7.1.11 For two i.i.d. r.v.s X1, X2, show that X X (t) = |X (t)|2, t !. (Hint:
1 2 1

Use Exercise 6.2.3 in Chapter 6.)


170 7 Stochastic Independence with Some Applications

7.1.12 Let X1, X2 be two r.v.s with joint and marginal ch.f.s X ,X , X and X . 1 1 1 2

Then X1, X2 are independent if and only if

X 1 ,X2
(t , t ) = (t ) (t ),
1 2 X1 1 X2 2 t1 , t 2 !.

By an example, show that

X 1 ,X2
(t, t ) = (t ) (t ), X1 X2 t ! ,

does not imply independence of X1, X2.

7.2 Proof of Lemma 2 and Related Results


We now proceed with the proof of Lemma 2.
PROOF OF LEMMA 2 Suppose that the r.v.s involved are continuous, so that
we use integrals. Replace integrals by sums in the discrete case. Thus,

k
( ) ( ) ( )

E g j X j = g1 x1 gk xk
j =1

) (
f X , , X x1 , , xk dx1 dxk
1 k

= g ( x ) g ( x ) f ( x ) f ( x )dx dx

1 1 k k X1 1 Xk k 1 k

(by independence)
= g ( x ) f ( x )dx g ( x ) f ( x )dx

1 1 X1 1 1 k k Xk k k

= E[ g ( X )] E[ g ( X )].
1 1 k k

Now suppose that the gjs are complex-valued, and for simplicity, set gj(Xj) = Yj
= Yj1 + Yj2, j = 1, . . . , k. For k = 2,

( ) [(
E Y1Y2 = E Y11 + iY12 Y21 + iY22 )( )]
( ) (
= E Y11Y21 Y12Y22 + iE Y11Y22 + Y12Y21)
= [ E(Y Y ) E(Y Y )] + i[ E(Y Y ) + E(Y Y )]
11 21 12 22 11 22 12 21

= [( EY )( EY ) ( EY )( EY )] + i[( EY )( EY ) ( EY )( EY )]
11 21 12 22 11 22 12 21

= ( EY + iEY )( EY + iEY ) = ( EY )( EY ).
11 12 21 22 1 2

Next, assume the result to be true for k = m and establish it for k = m + 1.


Indeed,
7.2 Independence:
7.1 Stochastic Proof of Lemma 2 andof
Criteria Related Results
Independence 171

( ) [(
E Y1 Ym+1 = E Y1 Ym Ym+1 ) ]
= E(Y Y )( EY ) (by the part just established)
1 m m+1

= ( EY ) ( EY )( EY ) (by the induction hypothesis).


1 m m+1

COROLLARY 1 The covariance of an r.v. X and of any other r.v. which is equal to a constant
c (with probability 1) is equal to 0; that is, Cov(X, c) = 0.
PROOF Cov(X, c) = E(cX) (Ec)(EX) = cEX cEX = 0.
COROLLARY 2 If the r.v.s X1 and X2 are independent, then they have covariance equal to 0,
provided their second moments are finite. In particular, if their variances are
also positive, then they are uncorrelated.
PROOF In fact,

( ) (
Cov X1 , X 2 = E X1 X 2 EX1 EX 2 ) ( )( )
= ( EX )( EX ) ( EX )( EX ) = 0,
1 2 1 2

by independence and Lemma 2.

The second assertion follows since (X, Y) = Cov(X, Y)/(X)(Y).


REMARK 5 The converse of the above corollary need not be true. Thus
uncorrelated r.v.s in general are not independent. (See, however, the corol-
lary to Theorem 1 after the proof of part (iii).)
COROLLARY 3 i) For any k r.v.s Xj, j = 1, . . . , k with finite second moments and variances
2j = 2(Xj), and any constants cj, j = 1, . . . , k, it holds:

k k
2 c j X j = c 2j j2 + ci c j Cov X i , X j
j =1 j =1 1i j k
( )
k
= c 2j j2 + 2
j =1

1i < j k
(
ci c j Cov X i , X j . )
ii) If also j > 0, j = 1, . . . , k, and ij = (Xi, Xj), i j, then:

k k
2 c j X j = c 2j j2 + ci c j i j ij
j =1 j =1 1i j k
k
= c 2j j2 + 2 ci c j i j ij .
j =1 1i < j k

In particular, if the r.v.s are independent or have pairwise covariances 0


(are pairwise uncorrelated), then:
iii) 2(kj= 1cjXj) = kj= 1c2j 2j, and
iii) 2(kj= 1Xj) = kj= 1 2j (Bienaym equality).
172 7 Stochastic Independence with Some Applications

PROOF
iii) Indeed,
2
k k k
c j X j = E c j X j E c j X j
2

j =1 j = 1 j =1
2
k
j = 1
(
= E c j X j EX j )

k
( ) + c c (X )( )
2
= E c 2j X j EX j i j i EX i X j EX j
j = 1 i j
k
= c 2j j2 +
j =1

1i j k
(
ci c j Cov X i , X j )
k
= c 2j j2 + 2
j =1

1i < j k
(
ci c j Cov X i , X j )
(since Cov( X , X ) = Cov( X , X )).
i j j i

This establishes part (i). Part (ii) follows by the fact that Cov(Xi, Xj) =
ijij = jiji.
iii) Here Cov (Xi, Xj) = 0, i j, either because of independence and Corollary
2, or ij = 0, in case j > 0, j = 1, . . . , k. Then the assertion follows from
either part (i) or part (ii), respectively.
iii) Follows from part (iii) for c1 = = ck = 1.

Exercises
7.2.1 For any k r.v.s Xj, j = 1, . . . , k for which E(Xj) = (finite) j = 1, . . . , k,
show that
k k

(X ) = (X ) ( ) ( )
2 2 2 2
j j X +k X = kS 2 + k X ,
j =1 j =1

where
2
1 k 1 k
j
X=
k j =1
X and S 2
= Xj X .
k j =1
( )